
Citation 
 Permanent Link:
 http://ufdc.ufl.edu/AA00034753/00001
Material Information
 Title:
 Modeling of transistor selfheating for circuit simulation
 Creator:
 Zweidinger, David T
 Publication Date:
 1997
 Language:
 English
 Physical Description:
 vi, 153 leaves : ill. ; 29 cm.
Subjects
 Subjects / Keywords:
 Electric current ( jstor )
Electric potential ( jstor ) Modeling ( jstor ) Parametric models ( jstor ) Simulations ( jstor ) Spices ( jstor ) Subroutines ( jstor ) Temperature dependence ( jstor ) Thermal resistance ( jstor ) Transistors ( jstor ) Dissertations, Academic  Electrical and Computer Engineering  UF ( lcsh ) Electrical and Computer Engineering thesis, Ph. D ( lcsh )
 Genre:
 bibliography ( marcgt )
theses ( marcgt ) nonfiction ( marcgt )
Notes
 Thesis:
 Thesis (Ph. D.)University of Florida, 1997.
 Bibliography:
 Includes bibliographical references (leaves 148152).
 Additional Physical Form:
 Also available online.
 General Note:
 Typescript.
 General Note:
 Vita.
 Statement of Responsibility:
 by David T. Zweidinger.
Record Information
 Source Institution:
 University of Florida
 Holding Location:
 University of Florida
 Rights Management:
 The University of Florida George A. Smathers Libraries respect the intellectual property rights of others and do not claim any copyright interest in this item. This item may be protected by copyright but is made available here under a claim of fair use (17 U.S.C. Â§107) for nonprofit research and educational purposes. Users of this work have responsibility for determining copyright status prior to reusing, publishing or reproducing this item for purposes other than what is allowed by fair use or other copyright exemptions. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder. The Smathers Libraries would like to learn more about this item and invite individuals or organizations to contact the RDS coordinator (ufdissertations@uflib.ufl.edu) with any additional information they can provide.
 Resource Identifier:
 028078824 ( ALEPH )
37857365 ( OCLC )

Downloads 
This item has the following downloads:

Full Text 
MODELING OF TRANSISTOR SELFHEATING
FOR CIRCUIT SIMULATION
By
DAVID T. ZWEIDINGER
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
1997
ACKNOWLEDGMENTS
I would like to thank Dr. Robert Fox for his patience, the Semiconductor Research Corporation for the financial support, the many people of Harris Semiconductor for their assistance and encouragement, Jon Brodsky for his long hours simulating heatflow, and finally my parents for doing it all before me and reminding me why it was worth the work.
TABLE OF CONTENTS
ACKNOWLEDGMENTS .......................................... ii
ABSTRA CT ..................................................... v
CHAPTERS
1 INTRODUCTION ............................
. . . . . . . . . . . . . . . . . .
2 BACKGROUND ............................................... 4
The SPICE GummelPoon Model .........
Thermal Impedance Modeling ...........
SelfHeating Models ...................
3 MODELING FOR CIRCUIT SIMULATION .......
An Overview of SPICE ...................
Formulation of the SelfHeating Model ....
Implicit Temperature Approach ..........
Thermal Impedance Modeling ........
The DC Solution ...................
AC Modifications ...................
Advantages and Limitations ..........
Explicit Temperature Node Approach .....
The DC Solution ...................
Implementation Options .............
SPICE Code Modification ............
Advantages and Limitations ..........
Convergence .........................
.. 4
~6 .... .10
.. ..... .. ..... ... 15
............... 15
............... 17
............... 18
............... 19
............... 20
............... 27
............... 28
............... 29
............... 29
............... 32
............... 34
............... 39
............... 40
4 SELFHEATING IN CIRCUITS .................................. 44
SelfHeating Effects .................................. 44
Simulation Examples .................................. 46
Convergence Comparison .............................. 57
5 MEASUREMENT OF MODEL PARAMETERS .................... 60
Measurement of the Thermal Impedance ................. 60
Temperature Coefficient Measurement ................ 62
Thermal Impedance Measurement .................... 67
Thermal Impedance Model and SPICE ................ 72
Selfheating and Model Parameter Extraction ............. 75
6 SELFHEATING FOR SOISPICE ................................ 83
SelfHeating Implementation ........................... 85
The SOI MOSFET Model ........................... 85
Specifying SelfHeating ............................. 86
Computing the Derivatives .......................... 87
Computing the Power Dissipation .................... 88
Local Temperature Update of Parameters .............. 88
Loading the Jacobian .............................. 90
Convergence Issues ................................ 92
Initial Simulation Results ............................. 93
Conclusion .......................................... 97
7 CONCLUSION ............................................... 98
APPEN DIX A ................................................. 101
APPENDIX B ................................................. 115
APPEN DIX C ................................................. 127
APPENDIX D ................................................. 140
BIBLIOGRAPHY .............................................. 148
BIOGRAPHICAL SKETCH ...................................... 153
Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
MODELING OF TRANSISTOR SELFHEATING FOR CIRCUIT SIMULATION
By
David T. Zweidinger
August 1997
Chairman: Robert M. Fox
Major Department: Electrical and Computer Engineering
Semiconductor devices consume power during operation. This power is dissipated mostly in the form of heat that increases the temperature of the device. This effect is known as selfheating. For bipolar transistors, BJTs, and silicononinsulator metaloxidesemiconductor fieldeffect transistors, SOI MOSFETs, selfheating can significantly change device performance. Circuit designers need an easy way to quantify the performance changes from selfheating, within the compact models used in today's circuit simulators.
A historical review of other attempts at modeling selfheating in circuit simulators is shown. Two approaches to including selfheating effects for bipolar transistors in the popular circuit simulator SPICE are presented. One approach uses an implicit temperature and transcendental equations to correct for selfheating in the device model. The other
approach uses an explicit temperature and an electrical subcircuit analogy for power generation and thermal dissipation to model selfheating. Techniques for assessing selfheating effects in hand calculations are shown, and example simulations with selfheating in common dc, ac, and transient circuits are shown. Reduced amplifier gain and thermal mismatching are common selfheating effects.
A new method for extracting the thermal impedance of bipolar transistors is derived, and results are shown for transistors of several different geometries. The effects of selfheating on the extraction of the GummelPoon model parameters are explored using longtime measurements and the modified SPICE in the optimization process.
The implementation of selfheating in the physicsbased SOISPICE model for SOI MOSFETs is shown. Common circuit examples are simulated showing the effects of selfheating. For analog circuits, the negative temperature coefficient of drain current can cause problems for biasing and amplifier gains. For digital circuits, selfheating is a function of frequency and is not significant.
CHAPTER 1
INTRODUCTION
In present circuit simulators, temperature is primarily modeled globally. Only a single temperature is used for all of the circuit elements, and the circuit is simulated isothermally at that temperature. This approach ignores selfheating, the local increase in temperature from the power dissipation of a device. The operating point of a device defines its power generation, and the geometries and materials affect how the power is dissipated as heat. The isothermal simplification is inaccurate considering that most circuits have many elements of different geometries, operating under different bias, and dissipating different amounts of power. The result is a performance change of the circuit because of all of elements operating at different temperatures. These performance changes from selfheating are significant for many circuits.
Semiconductor devices are especially sensitive to their operating temperature. One device that is particularly sensitive is the bipolar junction transistor (BJT). In BJTs the dominant sensitivity arises from the exponential dependence of the carrier concentration on temperature. Only a few degrees Kelvin of temperature change will cause significant changes in the carrier concentration. The BJT terminal currents are a function of the carrier concentration, and hence are greatly affected by
small temperature changes. Some temperature dependent mechanisms that are weaker than the carrier concentration are the carrier mobility and recombination lifetimes which become significant when large temperature changes occur on the order of tens of degrees Kelvin. These two mechanisms are important to silicononinsulator metaloxidesilicon fieldeffect transistors, SOI MOSFETs. Selfheating can cause failure of heterojunction bipolar transistors (HBT) where the large negative temperature coefficients cause currents to drop to zero in relatively lowpower operation [Gao89]. Including these effects in the compact BJT and SOI MOSFET models used in circuit simulation is important for improving accuracy in some circuits.
The importance of selfheating has become more evident with changes in technology. The continued shrinking of geometries, and the resulting increase in current densities and restriction of heat flow, tend to increase the thermal impedance and raise device operating temperatures. The thermal impedance is the temperature increase for a given power dissipation. The trend toward the use of silicon dioxide, Si02, for isolation makes selfheating more important, because the decreased thermal conductivity of Si02 compared to silicon creates a barrier to heat flow and greatly increases the thermal impedance of the device.
Finite elementbased simulators are available that can simulate a BJT under dc and ac conditions with selfheating [Lia94], but these are extremely slow for analysis and design of circuits with multiple devices.
What circuit designers need is a compact modeling simulator that includes selfheating. SPICE, with its GummelPoon model, is a popular circuit simulator. The temperature dependencies of all the parameters are implemented in the code for simulation of circuits at different global temperatures. The goal is to implement selfheating in the model by incorporating the temperature dependencies and the device thermal impedance and power dissipation to allow each device to be simulated at its own local temperature.
The following chapters are organized to give a understanding of the motivations for this work and the solutions reached for incorporating selfheating into circuit simulation. Chapter Two a historical view of thermal impedance modeling and selfheating simulation. Chapter Three describes two ways to incorporate selfheating into SPICE. Chapter Four focuses on circuit analysis in the presence of selfheating and shows some example simulations. Chapter Five shows a new extraction method for measuring thermal parameters of BJTs, in particular the thermal impedance. The effects of selfheating on model parameter extraction are also examined. Chapter Six shows the implementation of a selfheating model for SOI MOSFETs, with some simulation examples included. Finally, Chapter Seven summarizes the information and insights presented.
CHAPTER 2
BACKGROUND
This chapter presents an overview of previous work in modeling selfheating. First, the BJT device model equations used as the framework for selfheating simulation are reviewed. Following this is an investigation of thermal impedance modeling. Finally, an examination of previous attempts to model selfheating in BJTs will show the need for improvement.
The SPICE GummelPoon Model
The dominant model for use in simulating bipolar transistors in circuits was introduced in 1970 [Gum70]. This integral chargecontrol model was incorporated into the circuit simulator SPICE, developed at the University of California at Berkeley. Most circuit simulators in use today have a core based on SPICE and its GummelPoon BJT model. Therefore, SPICE was the obvious platform for the study of selfheating in BJTs and circuits. The communication of selfheating modeling is helped by the common use of SPICE based simulators in industry. What follows is a short overview of the SPICE GummelPoon model that will form the basis for most of the discussion in the chapters to come.
The following is the GummelPoon model for an npn BJT. The base current (IB) is composed of four current elements,
1B = IBF+IBE+IBR +BC, (2.1) where IBF is the forward diffusion current that results from backinjected holes across the baseemitter junction, given by
BE
is [NI.VT
IBF = e 1. (2.2)
where Is is the saturation current, I3F is the forward dc current gain, NF is the forward current emission coefficient, and VT is the thermal voltage. The nonideal baseemitter current results from recombination in the quasineutral base, and is given by:
V BE
IBE = ISE 1i (2.3)
where ISE is the baseemitter leakage saturation current and NE is the leakage emission coefficient. Similarly, the dual for the reversebias case exists. IBR is the reverse diffusion current given by: ( VBC
IBR = {eNT 1 (2.4) where PR is the reverse dc current gain and NR is the reverse current emission coefficient. Finally the nonideal basecollector current is
IBC =
(2.5)
where ISc is the basecollector leakage saturation current and NC is the leakage emission coefficient.
The collector current is given by
VBE VBC
IC = Ie e  IBR  IBC (2.6)
where QB is the base charge factor that accounts for basewidth modulation and highinjection conditions. The base charge factor is QB 1 + 4Q2) (2.7) QB 2 (( 12+
where
Q ( =1VBC VBE)1 VAF VAR)
and
V BE VBC ISE NFVT ' SC NRVT Q  1 +(2.9)
The parameters VAF and VAR are the forward and reverse early voltages, and IKF and IKR are the forward and reverse knee currents or corners for beta high current rolloff.
Thermal Impedance Modeling As a device heats from its power dissipation, its operating point shifts causing the power dissipation to change. This circular mechanism of heating is called thermal feedback. There are several thermal feedback
mechanisms in integrated circuits, which can usually be considered separately [Mei77, Fox91c]. One mechanism is the rise in the overall chip temperature due to the total power dissipated on the chip. This temperature rise is controlled by the chiptopackage and packagetoambient thermal impedances. This global heating operates over a long time scale (milliseconds to minutes), and couples all of the devices on a chip. Global heating can generally be reduced by careful packaging and heatsinking, and is accounted for in circuit simulation by simulating at multiple temperatures. Global heating is important to recognize in parameter extraction, where long thermal settling times cause the isothermal assumption to become inaccurate.
For largearea devices, or those dissipating large power, there can be direct coupling between the heat dissipated in one device and the temperature of other devices. This mechanism is strongly affected by the circuit layout. For example, thermal coupling in an operational amplifier from the output stage back to the input can profoundly affect the circuit's gain characteristics, although careful layout exploiting symmetry can mitigate these effects [Fuk76].
Direct heating of a transistor by its own power dissipation cannot be eliminated through packaging, scaling, or layout. This mechanism can cause substantial errors in modeling even without high power dissipation. As shown by Joy and Schlig, for modest power dissipation, the temperature rise is mostly confined within the transistor itself [Joy70].
Thus this effect can be simulated only by allowing each transistor to be modeled at its own temperature, controlled by its thermal spreading impedance and power dissipation. The thermal spreading impedance, which controls the temperature rise, can be predicted from the transistor's geometry [Lee92, Bro93, and Zwe95b].
The thermal impedance model derived by Fox and Lee [Fox9lb] is the frequencydomain equivalent of the timedomain derivation shown by Joy and Schlig [Joy70], and is based on the same assumptions. All of the heat is assumed to be generated uniformly in the rectangular volume of the collector spacecharge region (SCR). The analysis assumes an adiabatic plane at the surface of the transistor created by an image of the SCR heatsource placed above the surface. This approach assumes there is little heat lost through the top of the BJT and the heat is removed predominately by the conduction of the body of the semiconductor. For isolation technologies where SiO2 is used between devices, this may not be a good assumption [Dav92, Bro97a]. In the time domain, the response to a point impulse of heat, measured a distance r away in a uniform medium, is = __ ___ ( r2 "'
T(r, t) = Kt exp  (2.10) 8K(iT Kt)32 4Kt)
where K is the thermal conductivity and K is the thermal diffusivity. The Laplace transform of (2.10) gives the thermal impedance
ZTH(r,s)= 4Krexp(r4) (2.11) For s = 0, this equation gives the thermal resistance RTH for the point source.
An expression for RTH at any point r' = (x', y', z') in the semiconductor due to the heat generated in the collector SCR can be derived by integrating RTH over the collector SCR and its image source. For a rectangular vertical BJT the integral is
Rri$WLDHr') 1 w D+idxdydz +(D+H)dxdydz (2.12)
Rr"W'LD'Hr') 4nKf0; LD " r +_D r I (.2
where r = (x  x)2 + (y  y,)2 + (z  z')2, W and L are the width and length of the emitter, D is the collectorbase junction depth, and H is the SCR width. The result in (2.12) can be numerically approximated by [Fox9l a, Fox9lb]: RTII = 1 (2.13) where
fl(dh) = (0.058d+0.14)h+0.34d+0.28 , (2.14)
f2(a) = 0.98 + 0.043a  6.9 104a2 + 3.9 106a3 , (2.15) dD (2.16) W'
a E, a>1, (2.17) h = 2cB + ,bd)(qN J) (2.18)
NEPI is the epitaxial doping density in atoms/cm3, VCB is the collectorbase junction voltage and bi is the collectorjunction builtin potential. This expression was found experimentally to approximate RTH well for a wide
variety of rectangular emitter geometries [Fox9la].
Based on a comparison of (2.11) and (2.13), the same RTH could be computed by replacing the collector SCR with a point heat source placed a distance reff below the emitter, where reff can be defined as
1  2 ./WLf.f2. (2.19) Ceff  2TT K" R.1 2
This reff can be used in (2.11) to simply estimate for ZTH for all frequencies as
ZTHt= RTIï¿½ exp[reff(l +)J] (2.20) Equations (2.19) and (2.20) can also be used to model ZTH based on a measured value of RTH.
Figure 21 compares a normalized temperature response to the model given by Joy and Schlig [Joy70] and the Fourier transform of (2.20). Note that the pointsource representation overestimates the phase of ZTH. However, the magnitude of ZTH is predicted well, and the errors in phase coincide with small I ZTH I, so that the overall error is small. Note that I ZTH I varies over a much wider frequency range than would be predicted by a singlepole thermal equivalent circuit, in which I ZTHI would decay primarily over one decade of frequency.
SelfHeating Models
In 1964 Miller proposed modeling selfheating using an electrical subcircuit analog to the power dissipation through the thermal spreading impedance [M164]. As shown in Figure 22, the power generated in the
device flows as a current through an impedance, establishing a potential equivalent to the change in temperature. Only a single RC pole is shown,
250
200 From transform of (2.20)
 ... From [Joy70]
o 150
L. .
Q. 100
E
101
Time (s)
Figure 21
Normalized temperature response to a 1 W power step for a 7xlWOtm npn BJT, using the model in [Joy70] and the inverse Fourier transform of the frequency response of (2.20) [Fox93b].
Figure 22
Example of thermal subcircuit used to model selfheating.
but electrical networks representing multiple poles could be constructed to better model the distributed nature of the thermal spreading impedance [MWi164I.
An early attempt at computer simulation of selfheating effects was done by Latif and Bryant [Lat8l]. In this work a simple, single timeconstant, thermal subcircuit was implemented into the EbersMoll BJT model [Ebe54]. They used a linear network solver (WATAND) to simulate dc and transient effects of second breakdown in power BJTs. Later, this work was extended to include the more accurate GummelPoon model in the same solver for a study of selfheating effects on current mirrors [Mun9l ]. These two works failed to include the effects of selfheating in the ac analysis.
A singlepole thermal subcircuit was used to model temperature in SPICE by Vogelsong and Brzezinski [Vog89]. Results from a few circuit simulations were presented. Again, only dc and transient analysis were implemented, ignoring the important ac analysis. Their approach to computing the power dissipation included heat sources other than the basecollector junction which all have different thermal impedances that were not modeled. The algorithms they used to implement selfheating are vague.
In all of the above implementations, values used for the thermal impedance are empirical. The impedance value used is the one that makes the simulation fit the data best. This approach can neglect the coupling of
selfheating effects to other effects in the device behavior. Most of the implementations used only a singlepole network which poorly models the distributed nature of the thermal impedance. Finally, all of the simulators above require a voltage source in the thermal network to represent the ambient temperature. Repeating a simulation at a different ambient temperature would require modification of all of the thermal networks in the circuit.
The underlying physics of the heat flow either within the transistor or between transistors is often neglected in the above works. Heatflow analysis shows that except for rather highpower transistors, temperature gradients are mostly confined within the transistor itself [Joy70]. Usually adjacent devices are thermally coupled through the devicetochiptopackage spreading impedances, that have very long time constants, rather than devicetodevice impedances [Lee92]. The heat coupling effects can usually be ignored. Where the heating of adjacent devices cannot be ignored is in multifingered emitter BJTs and HBTs. The close proximity of each finger causes a strong thermal coupling. For example, for a three finger BJT driven with a constant current, the middle finger is heated by the two outside fingers causing its current to increase and the others to decrease. This effect can degrade the performance of multifingered emitter devices.
The improved implementation shown in the next chapter includes dc, transient, and ac analyses. It allows flexibility in formation of the
14
thermal impedance network, so that the self impedance and impedance to adjacent devices or emitter fingers can be modeled accurately.
CHAPTER 3
MODELING FOR CIRCUIT SIMULATION
Because of the deficiencies of previous attempts to include selfheating in circuit simulation, a more complete selfheating implementation was needed. The first priority was to create a simulator that could model the ac effects of selfheating. The frequency domain had been mostly ignored for analog circuits. The first implementation filled this void using an implicit local temperature approach. In the implicit temperature approach the BJT model is reduced to a twoport yparameter equivalent model. The yparameters are then modified by selfheating, and the modified yparameters are converted back to the original model form. Unfortunately, this implementation made transient analysis very difficult, so another implementation using an explicit temperature and thermal subcircuit was done. A thermal subcircuit represents the power dissipation, heat flow, and temperature rise. This requires a thermal impedance network with some inherent limitations, but does allow transient analysis, unlike the implicit node approach. Both of these approaches are described in this chapter.
An Overview of SPICE
To understand the implementations, it is important to understand first the organization and function of SPICE. The circuit simulation
program can be broken down into four pieces: preprocessing, dc and transient solution, ac analysis, and postprocessing.
The preprocessing section is responsible for reading the circuit description file and loading the information into the data structure. This processing includes building a instance matrix of all the circuit elements to preserve the connectivity information, and associating any parameter values and models that go with each instance. The data structure for the Jacobian matrix, voltage and current vectors, and state variables is then built. Any processing of model parameters is done, and the global temperature dependences are updated for the selected temperature of the simulation.
The dc and transient section of SPICE loads the Jacobian matrix with the circuit conductances and branch currents, and solves the resulting linear equations for the node voltages. For nonlinear elements an iterative NewtonRaphson method is employed. The device models are used in routines to compute the next guess for the conductances, currents, and voltages associated with each device. During transient analysis the energy of storage elements is integrated for each time step, and reduced to an equivalent conductance and current for the matrix load.
The ac analysis section uses the real part of the Jacobian from the solution in the dc analysis, and computes the complex part of the matrix for reactive elements at a given frequency. Perturbation of a node voltage
or branch current gives the smallsignal response at the linearized operating point.
Finally, the postprocessing routines gather the simulation results and prints or plots the data. The results can include device operating point information, sweep data from ac, dc, and transient analyses, and simulation statistics.
Formulation of the SelfHeating Model
The derivation of the both modeling approaches shown here starts the same way, with the expression for the smallsignal collector current. For the current into the collector including selfheating
alc ___1c = a V"Vbe + Vcb + L (3.1) Tfixed B T.fixed
where the power is
P = ICVce + icVCE + IBVbe + ibVBE. (3.2) Expanding the power term in (3.1) and reducing the partial derivatives to conductances results in
Ic T
c = goEVce+gmEvbe+ gEvcb+ (3.3) where the E subscript denotes electricalonly. Substitution of the thermal resistance gives
OIc
= gc RTHP. (3.4)
The last partial is substituted using the definition of the fractional temperature coefficient given by
Dm = (1/Im) dm (3.5)
where m is c or b for the collector or base current. The resulting substitution gives
ic = goEVce + gmEvbe + gtEVcb + ICDCRTHP. (3.6)
Equation (3.6) marks the point of divergence of the two implementation approaches. Solving for ic gives a ratio with temperature represented implicitly as done by MUller [Mti164]. This is the basis of the implicit temperature approach [Zwe92]. If the last term in (3.6) is represented using temperature for the thermal resistance and power product, a new matrix node is needed. This is the basis for the explicit temperature approach.
Implicit Temperature Approach
This is the simplest modification to make to the SPICE source code. The only changes needed are the addition of the new model parameters and modification of the device model routine. There are no new model matrix elements or state variables that would require changes to the device matrix. Detailed FORTRAN code modifications to SPICE2G.6 for this approach can be found in Appendix A.
Thermal Impedance Modeling
In this approach the thermal impedance is computed for each iteration of the device solution. New model parameters to describe the geometry of the BJT are added to implement the thermal impedance equations discussed in the last chapter. These parameters are: the length and width of the emitter, the depth of the collectorbase metallurgical junction, and the collector doping concentration. They are added to the list of values available on the BJT model line. From these four parameters and the collectorbase bias voltage the effective thermal resistance is computed from equations (2.13) through (2.18) for each iteration in the dc solution.
For the ac analysis, the effective heat source radius, reff, is computed from the thermal resistance computed in the dc solution or given in the model line. Then (2.20) is used to compute the thermal impedance at each frequency. The thermal impedance can also be computed from the thermal resistance and capacitance by addition of the thermal capacitance to the model line. This allows a simple singlepole model to describe the thermal impedance.
A hierarchy was formed to manage these three models. The first option is the lowest level where RTH and CTH (thermal capacitance) must be supplied and are used in the dc and ac analyses. The next option is used when only RTH is supplied. In this case, the given RTH is used in the dc solution, and ZTH is computed from reff for each frequency in the ac
analysis. The last option uses the geometry information and the predictive model to compute RTH for the dc solution and ZTH for the ac analysis. The DC Solution
The following is a brief description of the changes made to the routine that computes the dc solution for the BJT. Including the temperature dependence
A first step to all approaches to modeling selfheating is to copy all temperature dependences to the inside of the NewtonRaphson loop. Normally in SPICE the model parameters are updated for temperature once before each simulation is begun, allowing circuits to be sequentially simulated at different global temperatures. By moving the parameter dependences inside the NewtonRaphson loop, the temperaturedependent parameters for each transistor are updated for its local temperature in each iteration. The dependentparameter temperature functions are described below.
The thermal voltage is implicitly temperature dependent: VT = kT/q where k is Boltzman's constant and q is the electron charge. The temperature dependence of the dc current gain parameters stem from the dependence of the junction barrier height on temperature, and are given as
= : FO T7 (3.7)
\10/
and
T ,XTB(38
PR = R0(j (3.8) where To is the original model parameter temperature and XTB is the P temperature factor. The following current parameters dependences result mostly from the strong temperature dependence of the carrier concentration: is = IsO e factI, (3.9) fact In
ISEO NE
ISE XTB (3.10) (T/T0)
and
factIn
ISC0 N
ISC ( 0XTBe N (3.11) (T/T 0)X
The exponential factor in the above equations is
factln  EG I (3.12) VTO VT
where XTI is saturation current temperature factor and EG =EGO E T (3.13) E~~GAP T+TG
is the semiconductor energy gap. Values for the arguments in (3.13) are shown in Table 31 for common substrate materials. The SPICE BJT model assumes the energy gap is constant and uses a single parameter in the model line to assign a value. This has not changed for this implementation. Equation (3.13) is added to the temperature equations for the explicit node approach
Material EGO (eV) EGAP (eV/K) TGAP (K)
Silicon 1.16 7.02 x 104 1108 Germanium 0.67 4.56x10 4 210 Gallium Arsenide 1.52 5.41 x 104 204
Table 31 Energy gap temperature modeling
parameter values for common device materials.
to fix this deficiency in SPICE. New model parameters to assign values for the coefficients were also added to the model line. Computing the temperature
The temperature is not represented in the Jacobian in this approach. Instead, it is computed in each iteration from the thermal resistance and the power dissipation, T = P RTH. The power is computed in each iteration of the NewtonRaphson loop from the dissipation in the basecollector and baseemitter spacecharge regions, P = IcVC + IBVB. Note this is the real power dissipation, and does not include any reactive power that may be present in a transient simulation. The thermal resistance is either provided by the user or computed from the geometry using equations (2.13) to (2.18). Modifying the Jacobian
In the dc analyses the hybridt parameters are converted into twoport yparameters using
Y11E = g(+31
(3.14)
Y12E = g9 (3.15) Y21E = gmgg (3.16) Y22E =go+ g (3.17) where the hybridTi parameters are shown in Figure 31.
Next the BJT commonemitter yparameters are corrected to include thermal feedback using [Mil64] = mnE + DmRTHImIn (3.18)
1  DmRTHP
where m or n equal 1 for the base or 2 for the collector, YmnE is the uncorrected electrical yparameter, and Dm is the fractional temperature coefficient of the base or collector current. The collector current temperature coefficient in the forwardactive region of operation is estimated by [Zwe93]
base collector
t gmVbe go
emitter
Portl Port2 Y21
Yi I Y22 Y12
Figure 31 Conversion between standard hybridn and twoport yparameter models.
DC = f IIC I q 8/ (3.19) +FJ4IBE/IKF)
where
XTI + qEG/kT  qVBE/kTNF(3.20)
f= T
For the base current temperature coefficient in same region the expression is
f2 VBE XTB + Ibf fXTBII/(IBT) (3.21)
DB = BEIY  kTNE kbf(2 TNF
where
E~q
f2 = XTI + kT (3.22) The thermally corrected yparameters are then converted back to the hybridit form using the following ordered equations: g1 = Y12 (3.23) gn = YIIg9A (3.24) gm= Y21 + g (3.25) go = Y22  gt (3.26) The hybridn parameters are then loaded into the admittance matrix. Again, the advantage here is that no new matrix elements are needed in the Jacobian to compute the selfheating effects.
Flow control
The flowchart of the dc solution routine with selfheating is shown in Figure 32. First, the voltages from the Jacobian matrix solution are loaded, and the currents are linearly predicted using the conductances, the change in node voltages from the previous iteration, and the previous iteration currents. Using these predicted values of current and the node voltages, the power is computed, followed by the temperature using RTH. (If the predictive RTH model is used then the value for RTH comes from the previous iteration.)
Next, the convergence of the currents, voltages, and temperature is checked. The predicted currents and the node voltages are compared to the values of the previous iteration, and if each comparison is within its specified tolerance, the device has converged. If the convergence check succeeds then the statevariables are saved, the BJT is marked as converged and the model routine is exited. If the convergence criteria are not met then the predicted temperature is used to update the parameter temperature dependences and the currents are computed from the model equations described in Chapter 2.
With the currents computed from the model, the power, RTH, the temperature, and the fractional temperature coefficients are computed. The convergence is checked again for the new current values against the predicted values computed earlier. If the convergence check succeeds then the state variables are saved and the BJT is marked as converged. If it
Check convergence of predictions against
last iteration
Figure 32 Flowchart of the dc solution for the implicit node
selfheating approach.
fails the statevariables are saved and the matrix load begins. The BJT conductance values are transformed into yparameters using (3.14) to (3.17), corrected for selfheating using (3.18), and converted back to hybridn form using (3.23) to (3.26). The modified conductances are then loaded into the Jacobian. After all the BJTs in the circuit have been processed and loaded into the Jacobian, the linear equation solver is invoked to generate a new voltage solutions. SPICE continues to iterate until all the nonlinear devices converge, or a maximum iteration count is reached.
AC Modifications
For the ac analysis the smallsignal power is computed from (3.2). The hybridn model is transformed into complex yparameters as follows:
YHlE = 9n+g+j)(Cxcb C be+Cbc) (3.27) Y12E =  g jO+J(Cxcb+Cb) (3.28) Y21E = mj+ j(xgm  OCbc) (3.29) Y22E = g0 + g4 + jg D  Cbc (3.30) where Cbe and CbC are the baseemitter and basecollector capacitances respectively, xgm is the excess phase term, Cxcb is a baseemitter transcapacitor controlled by the basecollector voltage, and the other terms are standard hybridn parameters shown earlier.
The complex form of (3.18) where ZTH is substituted for RTH is used to modify the yparameters to include selfheating. The fractional
temperature coefficients used are from the dc solution. The thermal impedance is computed from RTH and CTH given by the user or predicted from the RTH computed in the dc solution.
The real parts of the yparameters are transformed back to the real parts of the hybridn model much as in the dc analysis. To transform the imaginary parts, first (3.30) is solved for Cbc. Then Cxcb and xgm are solved from (3.28) and (3.29). Finally, Cbe is solved from (3.27). After transforming the modified yparameters back to the original parameter set, the Jacobian is loaded and the matrix is solved for the current frequency.
Advantages and Limitations
This implementation is easy to use; with only a few new model parameters the user can model selfheating. The changes to the original SPICE source code are minimal as well, making the model easy to support. The bias and frequencydependent thermal impedance model makes for smooth, accurate dc and ac simulations.
The selfheating model in this implementation is only valid in the forwardactive region of operation. The most significant selfheating occurs in the forwardactive region, where the collectorbase junction has a large reversebias voltage and the high collector current create large power dissipation. The forwardactive region is also the most common region for analog applications where selfheating has the greatest effect on
circuit behavior. It is for these reasons that this implementation has focused on this region.
The greatest drawback to this implicit temperature approach is the inability to do transient simulations. SPICE is illequipped to solve the problem for this modeling approach, and this is the ultimate reason why the explicit thermal node method was implemented.
Explicit Temperature Node Approach
In the previous approach, the temperature was computed implicitly in each iteration using the voltage and currents of each iterate, and then modifying the existing internal conductances. In the following approach, temperature is represented explicitly by the addition of new elements to the device model, and a thermal subcircuit to model heat flow. A new row and column was added to the Jacobian along with new state variables for temperature and power dissipation. The dc, ac and transient analyses can be all implemented using this thermal node approach. Detailed FORTRAN code modifications to SPICE2G.6 for this approach can be found in Appendix B.
The DC Solution
The following is a brief description of the changes made to the routine that computes the dc solution for the BJT. Starting where the derivation of (3.6) left off, the new matrix elements are derived. If the substitutions of collector thermal transconductance, gTC  IcDc, and temperature,
30
T = RTHP, (3.31) are made in (3.6) then IC becomes
ic = goVce+gmVbe+gtVcb+gTCT (3.32) A similar approach used for the base current results in the expression ib = gnvbe  glVcb + gTBT (3.33) where gTB = IBDB is the base thermal transconductance. Using the electrical analogy for (3.31) results in modeling the temperature as a voltage and the power as a current. To include the power in the Jacobian, it must also be represented as a linear function of voltages and conductances. Substituting (3.32) and (3.33) into the expression for power in (3.2) gives the following:
p = (govce + gmVbe + ggVcb + gTcT)VcE (3.34)
+ (gvtbe  g11Vcb + gTBT)VBE + ICVce + IBVbe
Using the above equations (3.31) through (3.34) the intrinsic BJT element stamp is formed:
gm+ go + gp g gm  go gTC Vc ic
gt gt + gn gn gTB Vb ib (3.35)
gMgo gn (gm + go + gn) ( gTC  gTB) ve ic + lb
1 TI P
g pc  gpb RTH + Pt
where gPC = Vce(gm + go + gt) is the collector power term, g = Vbe(gt  gt) is the base power term, and gpt = VbegTB + VcegTc is the temperature power
term.
The device temperature change is computed at the new device node by a controlled current source equivalent to the device power dissipation in series with the effective thermal impedance as shown in Figure 33. The ambient or global temperature of the circuit is represented by ground, and the TA node in the figure is grounded for most cases.
The power dissipated in the transistor flows through the thermal impedance producing the effective temperature change of the device. The base and collector currents are then modified by the new thermal transconductance elements and the temperature. The thermal resistance is the sum of the resistances in the thermal impedance network, shown
B b c C
b
4 c 4
g_ 4TB AT gm Vbe go gTc AT
e
Re
E
ICTH1 ICTH2 CH3 AT 1 : ITA
tRVH1 RTH2 RH3
P
Figure 33 BJT internal largesignal circuit model showing
new temperature node and thermal impedance
ladder network.
here divided into three poles. Normally the thermal impedance network is connected to ground at the point TA, but it can be coupled to other devices through an additional network representing the impedance between BJTs. Implementation Options
Two choices affected the selection of an implementation strategy. The first choice was how to implement the thermal impedance subcircuit. The second choice was how to compute the temperature derivatives.
One way to implement the thermal subcircuit is similar to the implicit node approach above and would entail coding the subcircuit inside of SPICE. This requires the BJT matrix size to increase by a row and column for each new node in the impedance network. For maximum accuracy this could require as many as five new nodes, which would almost double the size of the BJT matrix. The topology of the impedance network would also have to be chosen in advance. The form of the ladder network shown in Figure 33 or the distributed form shown in Figure 34 could be used. Once implemented in the code, whichever topology is chosen cannot be changed by the user. The advantages of the internal approach are its RTH1 kH2 RTH3
AT AV
pCH TCTH3
Figure 34 Distributed form of the thermal impedance
subcircuit.
transparency to the user and the ability to use the biasdependent dc and ac impedance model implemented in the implicit node approach.
The other way to implement the thermal subcircuit is to leave it to the user to specify the thermal impedance network in the circuit description file just like any other circuit elements. With this approach the flexibility to use any topography is retained. Computing the thermal impedance is a task well suited for preprocessing using the SPICE input processing routines. The biggest advantage to this approach is that relatively few changes are needed to the BJT matrix and SPICE data structures, making the implementation much simpler.
The temperature derivatives can be computed analytically or numerically. When this work began the main interest was in selfheating effects on dc and ac simulations in the forwardactive region where the effects are greatest. An analytical solution of the temperature derivatives was practical when considering only the forwardactive region, and made for an elegant implementation in the implicit node approach. However, a problem arose when a device passed through operating regions where the derivatives were incorrect. The simple problem of implementing a single region became complex when analytical expressions for other regions had to be implemented. A numerical approach to computing the temperature derivatives improves the problem of regional boundaries. The numerical derivatives are computed using
Ax x (3.36) tT T T
where x is any temperature dependent variable at the temperature T and x' is the value of that variable at temperature T' a small increment above T. To implement the numerical derivatives requires each variable, in this case the collector and base currents, to be computed twice, once at each temperature. This takes about the same amount of computation as the direct analytical approach because of the complexity of the analytical expressions (see equations (3.19)(3.21)) compared to those of the currents. SPICE Code Modification
This section details the implementation. While many different approaches have been tried over the course of this work, only the final version is covered here. In this implementation the best tradeoffs for efficiency, flexibility, and convergence have been made. The thermal impedance is described externally in the SPICE input file, and the derivatives of the collector and base currents with respect to temperature are computed numerically.
Since the thermal impedance implemented externally to the BJT model, there are no new model parameters. Only one new device matrix node has been added for the temperature and four new state variables: temperature, power dissipation, and the derivatives of the collector and base currents with respect to temperature. The modifications to the SPICE data structure are relatively small, but are complicated to realize
because of the complexity of the SPICE data structure. Fifteen subroutines were modified and one new routine added to incorporate the changes needed for selfheating, but not all of these are discussed here for the sake of brevity. For more details, see Appendix B.
The format of the BJT device input line was changed to include the external thermal node. The new BJT line is as follows:
Qxxxx nc nb ne [ns [nt]] model [area] [initial conditions]
where the items in brackets are optional. The substrate node number (ns) and the temperature node number (nt) are both optional. The substrate node must always be given if the temperature node is used or differentiating between them when only one is specified would be impossible. An example circuit description with selfheating is shown in Figure 35 where the substrate node has been grounded. If the thermal node is not specified the node is assumed grounded and the standard isothermal model is used.
Once the preprocessing routines were modified to recognize the thermal node, build the new device matrix row and column, and add the new state variables, the work of modifying the device model routine for dc and transient analyses could begin. Note that the temperature state VCE 1 0 DC 5
VBE 2 0 DC 0.8
Q1 1 2 0 0 3 NPNMODEL
RTH 3 0 500
Figure 35 Sample modified BJT circuit description for
SPICE with a thermal resistance of 500 K/W.
variable stores the change in temperature above the circuit ambient, and is converted to the absolute temperature only where needed in the model. Unless otherwise stated, the temperature referred to in the following is the temperature change.
In the device initialization phase, the temperature is set to zero. No power has been computed so no temperature change can exist. Next in the linear extrapolation section the equation for predicting the power is added and temperature is added to the current prediction equations: P = P+ 11cAVcEJ + [IIBVBfl (3.37)
where P is the predicted power and P is the power from the previous iteration,
1c = I + (gm + go)AVBE  (go + gm)AVBc + gTCAT, (3.38) and
IB = IB + gnAVBE  gtAVBc + gTBAT. (3.39)
Next the initial convergence check is changed to include the temperature. The absolute tolerance on the temperature multiplies the conversion factor of 100 K/V and the tolerance of the junction voltages. With only a unity conversion factor, using the default voltage tolerances would require the temperature to be accurate to 50 tK for convergence. This is more precision than is needed and discourages convergence. Scaling the tolerance to 5 mK provides enough accuracy and improves the robustness of convergence. The power does not need to be included in the
check, because it is a linear function of currents and voltages and will converge when they do.
Similar to the implicit node approach, the temperature update equations were moved inside the model routine. The model parameters are adjusted for local changes in temperature in each iteration. After the initial convergence check and any junction voltage stepsize limiting, the temperature dependences are inserted. A loop is formed around the temperatureupdate equations, so they can be used in computing the numerical derivatives of the currents with respect to temperature. (See the flowchart in Figure 36.) The temperature in augmented by 0.1 K, the model parameters are adjusted for temperature, and the currents are computed from the model equations. At this point the hightemperature current values are temporarily stored and the 0.1 K is subtracted back from the temperature to return to the original value. The loop returns to the temperature update and then recomputes the currents for the proper temperature. The loop is exited and the current derivatives, gTC and gTB, are computed using the approximation of (3.36). The choice of 0.1 K for the temperature change was arrived at empirically; it results in enough change in the currents so the numerical problem of subtracting two close numbers of finite precision does not occur, and it is small enough to preserve accuracy in the approximation.
The power dissipation must be computed before any changes from reactive currents in the transient analysis can change the base and
Load voltages from matrix solution
Check convergence of predictions agains
last iteration
Correct parameters L Subtract delta for temperature H temperature
Flow chart of BJT model routine.
Check convergence of currents against
predictions
Figure 36
collector current value. The changes in current from reactive elements in the transient analysis are not real and do not contribute to real power dissipation. Such charging currents must be dissipated in the device parasitic resistances to generate temperature changes. Since the reactive elements of the thermal impedance are defined externally, no changes are needed to the transient analysis section of the routine.
SPICE includes a second convergence check after the transient section to compare the currents newly computed from the model to the predicted currents. As noted previously, it is not necessary to check the power; the temperature will not change within the model routine. The state variables are saved and the Jacobian loaded as shown in (3.34) and (3.35).
The changes to the ac analysis routine are minimal. In addition to the normal matrix load, the temperature row and column are loaded with the smallsignal equivalent of the dc matrix load. Advantages and Limitations
Many of the advantages and limitations to the external thermal impedance approach have already been mentioned, but is worth restating them. The advantages are many. Since the thermal impedance is implemented outside the BJT model, any topology or number of time constants can be used for the thermal equivalent circuit. No new model parameters and only a few new state variables are required; an internal thermal impedance would have many of each. The temperature
derivatives are computed numerically, improving continuity across operating region boundaries with little or no additional computational overhead. Transient analysis is possible where it was not in the implicit node approach. The biggest limitation is that the thermal impedance is fixed. Neither bias nor temperature dependences are implemented, and the smooth frequencydomain solution of (2.20) is unworkable in the smallsignal analysis. Versions of SPICE allowing nonlinear equations in userdefined controlled sources would make it possible to overcome this limitation.
Convergence
Selfheating makes convergence harder to reach. The most common problem with convergence is thermal runaway. SPICE sets the junction voltages to large values on the first iteration to ensure the devices are on. With the large forward biases the model equations compute large currents. After the first matrix solution, the returned node voltages tend to be quite large. These factors combine to overpredict the power dissipated in the device, causing an overprediction of the temperature. The overpredicted temperature in the second iteration can cause the temperaturedependent parameters to become large and cause the computed currents to increase enough to cause the temperature to increase even more. The temperature and current will grow geometrically until a numerical overflow occurs, usually after only a few iterations.
Sometimes a variation of this thermal runaway occurs where the simulated temperature becomes high in an iteration and the model overcompensates causing a negative temperature change, then positive again, oscillating with increasing magnitude with similar results as before.
Two controls have been put in place to minimize these problems: a hard limit that ensures that only positive temperature changes and a method in which the initial temperatureiteration is skipped. Theses enhancements were made to both implementations to improve the likelihood of convergence by preventing the thermal runaway condition.
For the iterationskipping method, the thermal resistance is always held to zero on the first iteration. This is done by checking the initialization flag and skipping the power computations when the flag is set. The zero power from the first iteration results in zero temperature change for the second iteration. The BJT currents and voltages usually settle into the solution region by this time, and the inclusion of selfheating is a much smaller perturbation on the solution. The resulting temperature computed from the power at this point will be closer to the actual value for the BJT, and subsequent iterations usually converge. The BJT model effectively executes two isothermal iterations before the selfheating is turned on.
To stop the oscillation a hard temperature limit was added at zero degrees local temperature change that prevents negative temperature
changes. If the model computes a negative temperature change at the beginning of any iteration the temperature change is set to zero for the remainder of that iteration. This allows the model to recover from the oscillation before it becomes a problem, and it prevents endothermic solutions not supported by the GummelPoon model.
A third method could be used to constrain runaway conditions in nonlinear equation solutions: stepsize limiting. This method limits the amount a solution element (for example: VBE) can change in a single iteration. This method can work for temperature in the selfheating model, but the best value to use for the maximum step size is difficult to determine and varies greatly from device to device and circuit to circuit. A step size that is too small can slow down convergence noticeably and one that is too large may only slow down runaway but not prevent it. For the BJT model this approach for convergence control gives little return for the computational investment compared to the other two implemented methods, so it is not used in this implementation. It could be useful, however, for special cases, especially if the user had control of the temperature stepsize for each transistor.
Results of convergence comparisons between the original isothermal model and the explicit thermal node approach are shown in the next chapter. Such comparisons provide only general estimates of the costs of using selfheating, since convergence in SPICE has seemingly random behavior for large circuits. For example, modifications to the input such as
43
changing the order in which the elements are listed can affect convergence.
CHAPTER 4
SELFHEATING IN CIRCUITS SelfHeating Effects
An easy way to understand selfheating effects in circuits is through the effects on BJT smallsignal parameters [Mtl164, Mtil70]. Assume that with selfheating neglected the commonemitter gparameters are given by
glE = 1/rnE (4.1) g12E 0 (4.2) g21E = gmE (4.3) g22E = I/roE (4.4)
Using the partial derivatives in Chapter 3, when selfheating is considered, the smallsignal collector current becomes
921EVbe + g22EVce + DCRTHIC(IBVbe + Icvce)
C =1  DCRTHP
assuming that ic/ib = ICAIB. Similarly, the base current is
lb = gllEVbe+ g2EVce+DBRTHIB(IBVbe + ICVce) (4.6)
1  DBRTHP
Thus the gparameters, corrected for selfheating, can be expressed as
SB DRTHI2 (4.7) gl2E + DBRTHIBIc
= 1  DBRTHP (4.7) g2 = 1DBRTHP (4.8)
g21E + DCRTHIBIC (49) _ g22E + D C (4.10)
21 =  DCRTHP 1  DcRTHP
The denominators in (4.7)  (4.10) become significant as the power approaches the critical values PC = 1/DCRTH or PB = 1/DBRTH where plots of IC and IB versus the port voltages have infinite slopes. These limits are related to the onset of second breakdown. In most analog circuits the power dissipation is kept much less than PC or PB, so that the effects of the denominator are negligible except in circuits requiring great precision. Note that as values for RTH rise, the power required to cause errors decreases.
Now consider g21. The electricalonly transconductance g21E is approximately gmE = IC/VT. If the denominator in (4.9) is close to unity, then the value of g21, corrected for selfheating, is approximately ï¿½2141 + DBRTHIBVT). Similarly, the electricalonly input conductance is 91E = 1/r.E = IB/VT, so that g1l = /rn g114 +DBRTHIBVT). For any reasonable conditions, the term DBRTHIBVT is very small, so these two parameters are only affected by the corrections in the denominators, which require substantial power dissipation approaching PBThe situation is different for the other two gparameters. A simple model for the electrical output conductance is 922E = l/roE = Ic/VA, where VA is the Early voltage. From (4.10), for P< PC) 922 = 922E( 0 + DCRTHICVA), where the second term in the parentheses can be thought of as a figure of merit giving the fractional error in the output
conductance caused by selfheating for low power. This term can be substantial even for very modest currents.
Neglecting selfheating, the effects of changes in VCE on IB, and thus g12E, are very small in forwardactive operation; typically g12E can be neglected in circuit analysis. However, including selfheating and assuming the denominator in (4.8) to be near unity, this parameter is much larger in magnitude and positive:
912  DBRTHIBIC = 22EDBRTHICVA/"O (4.11) This effect is exploited in the next chapter to extract the thermal impedance of the BJT.
Simulation Examples
The repercussions of the above results for circuit designers are evident in a variety of circuit types. The following examines the effects of selfheating on some common circuits. The SPICE circuit descriptions for all the circuits shown are provided in Appendix D. All the simulations used the explicit thermal node simulator with thermal impedance subcircuits.
The most common analog circuit is the commonemitter amplifier configuration where the emitter and base impedances are typically small and output (collector) load is large, comparable to ro. Here selfheating can significantly raise the input resistance and decrease the output resistance. The result is a greatly reduced gain at low to moderate frequencies where selfheating predominates. This effect is illustrated in Figure 41
[Fox93b]. This commonemitter amplifier circuit was originally published by Lee using the implicit temperature simulator developed in this work [Lee92]. The rest of the circuits are new examples of selfheating effects on circuits.
The larger 741 opamp circuit shown in Figure 42 is a common twostage amplifier circuit with a low current, low gain differential input stage and a high current, high gain stage [Gra84]. The dc transfer
75
70 'S 65
a
60 55
50
102
103 104 105 106 107 108
Frequency (Hz)
Figure 41 Simulated voltage gain vs. frequency for
commonemitter amplifier with and without selfheating. Note that the dc gain is halved by selfheating.
characteristics of a 741 opamp circuit are shown in Figure 43 both with and without selfheating. The selfheating curve shows a significant nonlinear distortion. This distortion causes a 10% shift in offset voltage and a large reduction in gain. The greater distortion at low output voltage
Figure 42 Schematic for the 741 operational amplifier
circuit. Resistance values are as given by Grey
and Meyer [Gra841
49
indicates the PNP currentmirror in the second amplifying stage, which has no emitter degeneration, has a much reduced output resistance due to selfheating.
For the 741 opamp, the first stage is operated at such a low current that selfheating effects are minimal. The second stage is operated at higher currents, and is driven from the low output impedance of an emitter follower, so there is an error in the dc gain due to selfheating. This is shown in Figure 44. When Miller compensation is applied around the
0 1 2 3 Input Voltage (mV)
Figure 43 Simulation of dc transfer characteristics of a 741
opamp with and without selfheating.
second stage, the dominant pole frequency is also shifted by selfheating. However, the negative feedback through the compensation capacitor swamps out the positive thermal feedback above the dominant pole frequency, so there is little thermal effect on the phase margin.
Some largesignal circuits where precise VBE matching is important can be affected by selfheating when the collectoremitter voltages of the matched BJTs differ. The BJT with the higher VCE has greater power
100
80
60 4..
40
0
10)8
Frequency (Hz)
Figure 44 Simulation of a 741 opamp openloop voltage
gain with and without selfheating showing the
effects of compensation.
51
dissipation, causing its VBE to decrease to suppress the increase in collector current. These dc effects were demonstrated for current mirrors by Munro and Ye [Mun9l], and for current reference circuits by Fox and Lee [Fox93b]. When these circuits are used in switching applications the VBE shifts settle over long time periods as the BJTs heat and cool.
Consider the pindriver circuit shown in Figure 45. The function of this circuit is to transfer a voltage level at lowcurrent, high impedance circuit to a highcurrent, lowimpedance circuit (or package pin). Going from VIN to VOUT the signal increases two diode drops through Q1 and Q2
Vcc
QQ4
Q2__j OUT
VIN 1
FpVEE
Figure 45 Simplified pindriver circuit.
then decreases two diode drops through Q4 and Q2 If all of the transistors are matched the output follows the input.
Transistors Q3 and Q2 are diodeconnected and do not dissipate much power. Q1 and Q4 can have large collectorbase voltages and can dissipate significant power. Figure 46 shows the temperature transients of Q1 and Q4. For this simulation, the supply voltages VCC and VEE were 10 V and 10 V, the current sources I, and 12 were 1 mA and 10 mA, and
40 30 20
04
0 50 100 150 200 Time (s)
Figure 46 Simulated temperature transients of Q, and Q4 of
the pindriver circuit for and 18 V input pulse.
250
the input voltage pulse switched from 2 V to 18 V. The temperature swing of Q4 is much larger than that of Q1.
The output voltage transients are shown in Figure 47 with and without selfheating. The long thermal transient can be seen in the long output voltage settling time. While the 0.1% settling time for the isothermal case is less than 10 ns, with selfheating the settling time is about 100 ps. The long thermal tail is a problem for highspeed precision applications.
8.05
8.00 7.95
7.90 7.85
7.80 7.75 7.70
100
150
200
250
Time (s)
Figure 47 Simulated transient output voltage for the pindriver circuit with and without selfheating.
. . . ï¿½  ï¿½ ï¿½ ï¿½ ï¿½ i thou S . . . . . . .i.. . . . . . .
 Without Selfheating SWith Selfheating ....... Input Pulse
ï¿½ , I ï¿½ . I . . I , *
Another application of selfheating in circuit design is thermal coupling of transistors that are near each other. The distance for a device to be considered near to another in thermal terms depends on the circuit. Fukahori and Grey showed that for a operational amplifier the input transistors could be affected by the heat generated in highcurrent output transistors if they are located close to the input transistors [Fuk76]. They showed that careful layout that located the output stage away from the input stage could cure this problem. The general rule in these cases is to lay out highpower elements of the circuit away from the sensitive subcircuits of a design.
For the case of multipleemitter BJTs the above rule does not apply. These devices are used in highfrequency applications primarily because of the reduced parasitics from the shorter intrinsic base region and reduced current crowding of the thinner emitter fingers. To simulate a multiemitter BJT with selfheating, it can be separated into multiple devices, one for each finger. This method was developed in collaboration with Brodsky [Bro97b]. A schematic for a threefinger BJT is shown in Figure 48. The fingers are thermally coupled by currentcontrolled voltage sources (Hsources in SPICE). Each finger's thermal subcircuit includes an Hsource for each of the other fingers. An Hsource is controlled by the power flowing in another thermal subcircuit multiplied by the thermal resistance between the two fingers.
Figure 49 shows the result of a dc simulation of the threefinger BJT. The middle finger is heated more than its neighbors because of its close proximity to both of the side fingers. It carries more of the base current and contributes more to the total collector current. The result is a reduced output impedance over the nonthermal simulation of about 70% for this example.
)Q1
Q2I
+
I 
HTH1
HTH1
HTH3
Figure 48 Schematic for simulating a threefingeremitter
BJT, and the three coupled thermal subcircuits.
I I
Q3
The simulation of circuits with selfheating adds to the accuracy of many designs, but is not necessary for all circuits. To further increase the reliability of simulation results, more accurate model parameters are needed, including thermal impedance parameters. A technique for extracting the thermal impedance for BJTs is introduced in the next chapter. The effects of selfheating on the GummelPoon parameter extraction are also examined.
15.0
10.0
0
o 5.0
0j
0.0 '0.0
2.0 4.0 6.0 8.0 10.0
CollectorEmitter Voltage (V)
Figure 49 Contribution of the middle and side fingers to the
total collector current of a threefinger emitter
BJT.
Convergence Comparison
The simulation statistics for each of the above circuits were collected for the selfheating and isothermal cases. The results of a three run average are shown in Table 41. The statistics for the multifinger
Table 41 Circuit simulation statistics.
Operatingpoint Ac Transient
Circuit Normal/Selfheating Normal/Selfheating Normal/Selfheating
number/time (s) number/time (s) number/time (s) Common 5/0.03 6/0.03 46/0.07 46/0.12 emitter NPN
741 Opamp 15/0.45 15/0.95 46/1.08 46/1.92 Pindriver 8/0.07 9/0.12 710/2.70 856/5.85
transistor are omitted, because there is not an isothermal equivalent circuit for comparison. These statistics are reported by the ACCT option in SPICE the reports on execution times and iterations for each analysis. The circuits were simulated on a Sun Microsystems, SPARCstation 10 with 64 MB of memory. The computer, at the time of simulation, was only running the FVWM window manager and common UNIX daemons, so the simulation time for each circuit was relatively unaffected by resource demands by other processes. The quality of the computer isolation was shown to be effective, since each of the averaged simulations returned the same result.
It is clear that simulations using selfheating have a greater cost in computation time. For the case of the singleBJT commonemitter circuit difference in the operatingpoint solution times is lost in the accuracy of
the measurement, but the selfheating circuit needed an extra iteration implying the need for greater computation time. The ac analysis used 71 percent more time because of the added elements of the thermal impedance subcircuit. Because of the linear nature of the ac analysis, the number of iterations is always exactly the same as the number frequency points computed.
For the 741 opamp, the iteration count for the operatingpoint solution is surprisingly the same, but there is a 111 percent increase in computation time for selfheating. It is important not to infer too much from these results, because the bias used for each case was different due to the shift in operatingpoint caused by selfheating. A small bias change for this class of circuits will almost always have an effect on convergence independent of selfheating. The additional time is due to not only the added circuit elements but the increase computation need in the BJT model for selfheating. The ac solution took 78 percent longer, which is consistent with the commonemitter circuit results.
The pindriver circuit's initial operatingpoint solution required one more iteration and 71 percent greater time per iteration for selfheating. The effects of selfheating on transient iterations and computation time are expressed in the timestep control. The SPICE algorithm to compute the timestep in the transient analysis increases the size of the time step each iteration until convergence is not reached and then "backs up" to try a smaller timestep. For the isothermal simulation, there are practically
no changes in the output for the majority of the analysis time. SPICE can take large timesteps and complete the simulation with fewer iterations. For the selfheating case the long thermal tail requires smaller timesteps and a greater number of iterations. The time per iteration increased 80 percent.
To keep the added overhead of selfheating to a minimum, it is important to use the insight provided in the first part of this chapter to choose the devices most affected for simulating with selfheating. This mixed approach also can improve convergence behavior over using selfheating for every BJT in the circuit.
This chapter has used models for thermal impedance for simulation. In the next chapter, the technique used for extraction of the thermal impedance is presented. It was shown here that selfheating changes device performance. The next chapter shows how selfheating can affect the extraction of the GummelPoon model parameters.
CHAPTER 5
MEASUREMENT OF MODEL PARAMETERS
This chapter describes methods for extracting model parameters for use in selfheating circuit simulation. First, a twopart, dc and ac method for extracting the thermal spreading impedance is shown. The results can be used to either verify an impedance model or to directly extract the impedance network for each device [Zwe95a, Zwe96]. Second, the effects of selfheating on GummelPoon dc model parameter extraction are examined. Selfheating causes errors in certain highcurrent parameters if they are extracted under nominally isothermal conditions unless selfheating is accounted for in the optimization phase [Dav88].
Measurement of the Thermal Impedance
Two components must be characterized for an accurate model: the thermal spreading impedance (temperature response to a change in dissipated power), and the temperature dependence of the currentvoltage relationships of the transistor. At first glance measuring the temperature dependence of the currents appears straightforward; the ambient temperature is varied by a known amount and the resulting changes in the terminal currents are measured. Actually the situation is more complicated because the actual device temperature rises above the ambient by an unknown amount as a result of selfheating. This unknown
temperature change can be significant compared to the ambient temperature change, especially in highcurrent regions. This chapter describes a way to overcome this problem, and a way to use the results to extract the thermal impedance from transient response.
Measurement of the base current response is treated as the thermometer in the techniques described here. An underlying assumption of these measurements is that collectorbase conductance from the modulation of the base recombination current is negligible compared to the effect of selfheating on the total base current. If operation is in the forwardactive region where significant impactionization can be avoided, any change in base current resulting from a change in collectorbase voltage with constant baseemitter voltage is dominated by selfheating. Characterization of the thermal impedance requires two steps. The first step involves finding the fractional temperature coefficients: the fractional change in base or collector current for a unit change in the transistor temperature with fixed baseemitter voltage. We find this information by sweeping the dc collector voltage of the transistor at different ambient temperatures and then cancelling the selfheating. The second step is to find the thermal impedance from the base current response to a collector voltage step. Using the temperature coefficient results from the first part, the basecurrent response is converted to a temperature response. After normalizing the transient temperature response by the dissipated power, the thermal impedance model can be fitted to the waveform.
Temperature Coefficient Measurement
In this part of the measurement cycle the goal is to calibrate the basecurrent thermometer by measuring the change in current caused by a change in temperature with fixed base voltage. When the temperature is increased the collector current and the power increase, causing the local temperature to increase considerably because of selfheating. An error arises whenever selfheating contributes significantly to the device temperature (at moderate to high current levels). Transient measurements, to be discussed in the next section, are typically made in these regions, and there needs to be a way to characterize the base current dependence on temperature in the presence of selfheating. The total temperature change of a BJT, including selfheating, is given as
AT = RTHAP + ATA = RTH(ICAVC + AIcVc) + ATA (6.1) where AT is the total temperature change at the emitter, ATA is the ambient temperature change, and RTH is the thermal resistance. Separately measuring the change in base current caused by a change in ambient temperature and then by a change in power gives two equations that allow cancellation of the selfheating from the temperature coefficients. The power is controlled by varying the collector voltage in the forwardactive region where basecollector conductance is negligible and avoiding regions where impact ionization is significant.
When changing the ambient temperature with the collector voltage, Vc, constant, (6.1) gives a net temperature change of
ATAvc 0 = ATA + RTHVCAIC (5.2) where
8IC
AIC =y AT = DcIcAT (5.3) is the change in collector current due to a change in temperature. Substituting into (5.2) and solving for AT gives ATA
ATh~v_ 0 1  DCRTHICVC When TA is held constant and VC is varied, (5.1) reduces to
ATIATA = 0 = RTH(ICAVC + AIcVc) (5.5) where
_IC _IC AVC AIc = AT+ AAV = DcIAT+ C (5.6)
C T avc C VA
and VA is the Early voltage. Substituting into (5.5) and solving for AT gives
IcRTH 1 + VC)AVc
ATIATA = 1  DCRTHICVC
Notice that the denominators of(5.4) and (5.7) are the same. Combining (3), (5) and the basecurrent dependence given by
l IB A IB BIB(AT8
B AT+ AVC =DBIATA + RTHVCAIC) (5.8) where the second term is the basecollector conductance (assumed to be negligible), results in the following ratio:
(AIB/AVc)I = 0
(l/T = ICR TH (I + Vc/VA). (5.9) (AIB/ATA) IV = 0
This can be solved for RTH if VA is known or is much greater than VC. It is important to note that the thermal resistance found here is an effective value, and includes the thermal impedance of the package and die. While this is not an exact value for the intrinsic thermal resistance, it is a good estimate for extremely low frequencies.
Once RTH is known, (5.8) can be solved for the fractional temperature coefficient of base current: D AIMB/IB (.0
DB =ATA + RTHVcAIC AVc = (5.10)
Note that this rather complicated procedure is not needed at moderate to low currents, where selfheating does not need to be cancelled.
The following is a description of the measurement setup and algorithm. The transistor is connected in commonemitter configuration inside an oven or on a probe station thermal chuck. It is important to have stable, accurate ambient temperature control during the measurement, because the voltage sweep takes a long time and errors in the ambient temperature significantly affect the accuracy of the results.
The algorithm for the measurement is simple, though very timeconsuming. First, set the ambient temperature and allow it to settle completely, chose a base voltage for the region of interest (moderate to
high current region), and sweep the collector voltage in the forwardactive region where a noticeable, linear increase in IB can be seen (avoiding any falloff due to impactionization). Use a very long delay time between measured points; about 3060 s typically works well. This allows the selfheating of the transistor to settle before making the measurement at each point. Execute the sweep, measuring IC, IB, and VCE for each VBE of interest. Change the temperature and repeat the voltage sweeps done before (from experience, a temperature step of less than 10 0C is recommended to avoid errors caused by the strongly nonlinear dependence of current on temperature). Take the average current changes with respect to collector voltage and ambient temperature. Using a geometric mean helps compensate for the nonlinearity of the data. Use (5.9) and (5.10) above to compute the fractional temperature coefficients at each base voltage value.
Figure 51 shows a plot of the temperature coefficients versus base voltage for an NPN Harris UHF transistor with a 3x50 gim2 emitter. The UHF process yields highfrequency, dielectrically isolated, vertical transistors [Dav92]. The collector voltage was swept from 24 V while the base voltage was varied from 700850 mV in 5 mV intervals using a HewletPackard 4145A Semiconductor Parameter Analyzer. External computer control was used to run the HP 4145A, because only short durations are available for the voltage steps using the integrated controller. The automation of the experiment also helped reduce human
errors and total measurement time by integrating setup, measurement, and data collection into a single routine. The ambient temperature was controlled using a Delta Design Model 111 Thermal Chamber; temperatures used were 25 and 33 'C.
This measurement has errors that are dominated by oventemperature accuracy, nonlinearity of the data, and effects of base and emitter resistances. The thermal chamber has a published repeatability
. 0.060
00.040 E 0.030
0.020 '
0.70
0.75 0.80
Base Voltage (V)
Figure 51 Harris 3x50tm UHFN3 bipolar transistor
fractional temperature coefficients of base and collector currents versus baseemitter voltage measured with and without correcting for selfheating.
0.85
accuracy of 0.5 0C and a relative error of 0.1 'C. The absolute error is not as important for the temperature coefficient measurement as the relative error since the change in temperature is used in the computations. Because the measurements shown here were repeatable with an error consistent with the 0.1 'C, the published errors can be believed. The emitter and base resistances cause a slight error in the intrinsic VBE for high currents. When AIC is 1 mA the intrinsic AVBE is reduced by 2 mV. The total error for the experiment is approximately 10 percent. Thermal Impedance Measurement
The objective of this part of the measurement cycle is to measure the normalized thermal response to a step in power. This is done by converting the base current response into a temperature response using the temperature coefficients extracted above. Solving the fractional temperature coefficient definition for the change in temperature gives IB(t) IB0
AT(t) = (5.11) IBDB
where IB is the median value of the base current for the transient. At this point the signal represents the thermal transient for the power step applied, and it must be normalized for a unit power step. This is done by dividing the AT(t) signal by the power of the step:
ATnorm(t) = AT(t) (5.12) A1cVc + AVcIc
where the denominator term is the magnitude of the power step.
The setup for this measurement is similar to the temperature coefficient commonemitter configuration with a collector voltage step applied instead of a slow voltage sweep. The setup is shown in Figure 52, where the BJT in the center is the test device being probed on a wafer. The currenttovoltage converter transforms the base current into a voltage waveform for the oscilloscope. With the base voltage held constant, the base current is
IB(t) = (VoUT VBE)/RF. (5.13) where RF is the feedback resistance. Unfortunately, the intrinsic baseemitter voltage is not constant because of the parasitic base and emitter resistances. Since the change is small compared to the thermal voltage, a linear correction is suitable:
IB(t) = RVOUT  VBE
RFRB(1 +3)RE (5.14) where RB and RE are estimated base and emitter resistances and P is the largesignal current gain.
RS=50K2i RF=5 10Q
VC Pulse
Generator '
0 +
 i VB
Figure 52 Schematic for the transient response
measurement.
There are important practical tradeoffs in choosing the feedback resistor value. The most important is the gain versus bandwidth tradeoff. Gain must be large because the signals of interest are typically very small, yet a wide bandwidth of about 100 MHz must maintained to resolve the earliest part of the transient response. In our measurements, an amplifier with gainbandwidth of 1 GHz was used with RF = 510 Q and RSCOPE = 50Q for a loop gain of about 10 and a bandwidth of 100 MHz. The "sponges" shown in Figure 52 are highfrequency sinks that absorb the energy at frequencies beyond those of interest in the thermal measurement; they are required to keep the UHF transistors from oscillating. The sponges also help absorb some of reflected power from the unmatched impedances of the pulse generator and BJT collector. The device under test was probed onwafer using one meter, 50 ! coaxial cables.
The following is a description of the method used to make the stepresponse measurement. First, a value of base voltage was selected to give a current large enough that the collector voltage step created enough power to get a measurable change in base current for a square wave collector voltage of 25 V. The best frequency to use for the square wave depends on the transistor; the response must settle before the end of each half cycle. Much of the response occurs in the first 50 p s, but it continues until about 1 ms, after which package heating dominates. Because of the distributed nature of the response, the opamp output was recorded using three overlapping time windows of increasing size to capture with high
resolution the fast beginning of the response while keeping the number of data points to a minimum for the long tail. Next, the pulse generator was set to dc output and an ammeter was connected in series with the collector to measure the collector current at the pulse maximum and minimum values. The opamp output is also measured at this time to compute the base current from (5.14). From this data the P of the transistor is determined, so that the transient base current can be converted into a collector current for computation of the power step magnitude. The base current minimum and maximum in this dc measurement differ from the values in the transient because of the difference in mean power being dissipated. Using (5.11) and (5.12), the normalized temperature response can be computed. Figure 53 shows the normalized temperature responses for a range of geometries.
Figure 54 shows the thermal resistance of the same transistors versus emitter width. Estimated errors for the measurement are also shown. These errors are primarily from three sources: errors in the extracted fractional temperature coefficient, the variance of P over the range of collector voltage, and the absolute error of the measured base current.
The absolute error of the base current seriously limits accuracy for small transistors that operate at low currents where it is difficult to get a large enough change in IB to be seen over the noise of the measurement equipment. This does not mean that thermal effects in these transistors
are insignificant. In the setup used here the limit for de collector current is approximately 1 mA. For a small transistor biased at 1 mA undergoing a 5 mW power step with an RTH of 500 K/W, an approximately 15% increase in IB occurs. However, this change is seen as only 1 mV at the opamp output. This signal is hard to resolve with a oscilloscope with a 1 mV/ division minimum resolution. Add to that a typical 2 mV of noise superimposed on this signal, and it becomes very difficult to resolve the transient even with averaging.
400 1 . . . . .  . . . . . .  I . . . .  ï¿½ M
350 300 250 200 150 100
105
Time [s]
Figure 53 Normalized temperature response versus time of
the UHF transistor for seven different emitter
widths.
10"3
Thermal Impedance Model and SPICE
The above results must be compressed into a manageable model for use in circuit simulation. Using the version of SPICE modified to include a temperature (AT) terminal described in the previous chapter, the thermal impedance can be incorporated into precision circuit simulation of dc, ac and transient. To convert the measured stepresponse data to an equivalent RCladder model, the program TIPP was used to provide SPICE with calculated values for the thermal resistance and thermal capacitance components needed for the thermal equivalent circuit [Bro93]. TIPP can handle up to five poles for the thermalimpedance model, and input can come from either measured transient data (AT vs. time as found
400 . . . .
,300
ï¿½ 200
H100
0 I . I
0 50 100 150 200 250 Emitter Width [gm]
Figure 54 Thermal resistance versus emitter width for the
UHF transistor with emitter length of 3 jim
here) or a predictive physical model [Joy70]. Figure 55 shows the normalized temperature response for a 3x70 pm2 transistor and the resulting threepole and singlepole fits. The threepole fit is good, and generally is the best tradeoff between accuracy and complexity for precision applications. For some situations the singlepole model is sufficient.
Many issues must be considered when deciding whether to use measurement or a physicsbased predictive model for determining the thermal impedance in a process. Typically, manufacturers extract model parameters for each transistor in a process from measurement. The
200.0 Measured Data .. . .
 Threepole Fit
1500..... Onepole Fit
150.0
o 100.0
. 50.0 ,
Z 0.0  "
50.0 . .
50.0.. .... .... . ...... . .., , . . ..
107 106 10o5 104 103 Time [s]
Figure 55 Comparison of normalized temperature response
for one and two pole fits and measured data of the
UHF 3x70 pim transistor
extraction is automated and can be done fairly quickly. If the measurement of the thermal impedance is included in the extraction, the time spent on each transistor is greatly increased.
Consider just the calibration step of extracting the current temperature coefficients. Estimating that each data point takes 30 s and that the minimal of two VCE values per VBE (which could case large errors) for two temperatures, gives a minimum of 2 minutes per VBE. Add to that the time for the change of temperature of approximately 10 min and the time to find the correct operating region for the measurement. This takes about 30 min per transistor before the transient measurement has been made.
Considering the time that would be required to measure parameters for every BJT geometry, development of a model for thermal impedance dependent on process parameters that scales with transistor geometry would be preferable. Such a model should use physics as much as possible but exclusively. The tradeoff here is the accuracy. A physical model cannot include all effects and still yield results quickly. For each neglected effect the accuracy usually decreases. For some structures the assumptions needed to solve the heatflow equations may not allow enough accuracy to be useful. For these structures the loss of accuracy in the physics can usually be corrected by using empiricism and the information from measuring only a few transistors. In the case of a vertical, junctionisolated, rectangular emitter BJT good predictive physicsbased models
exist (such as the one described in Chapter 2). For the waferbonded UHF transistors measured here, a scalable thermal model is difficult to develop due to the large process variation of the oxide depth that greatly affects the thermal impedance. For thin SOI technologies where the isolation depth and thickness are wellcontrolled, physical modeling seems simple until the increased importance of heatflow out the contacts and wires is considered. Extraction of thermal impedance for some transistors in a process will always be needed either to validate a model or to tune some empirical parameters. A reasonable goal is to avoid spending needlessly long times measuring every BJT in a process.
Selfheating and Model Parameter Extraction
When standard parameterextraction procedures are used to find parameters such as Early voltage and knee current, the data typically include some selfheating effects. When these parameters are used in a model with selfheating, the selfheating effects are overrepresented. For accurate modeling, the parameter extraction procedures themselves must be modified to separate thermal and electrical effects. This section describes a process to accurately extract these highcurrent parameters.
Most of the SPICE GummelPoon model parameters are normally extracted in lowpower regions where selfheating effects can be ignored. Three parameters, however, are usually extracted in highpower regions where selfheating can be important. These are the knee current, the Early voltage, and the parasitic base resistance (RB). IKF and RB are
extracted from a Gummel plot at high current, usually with a significant basecollector voltage to avoid quasisaturation, so that power dissipation may not be negligible. VA is extracted from the slope of a collector current versus collectoremitter voltage plot, which is very sensitive to the power dissipation of the measured device. It is tempting to try pulse measurements to avoid selfheatinginduced problems by measuring the device before it can heat up. Unfortunately, because the thermal response in scaled devices has time constants on the order of nanoseconds, it is very difficult to measure a BJT in electrical steadystate before significant selfheating begins. In the approach presented here, the device is allowed to reach thermal steady state before data is recorded.
For comparison, a normal isothermal extraction not accounting for selfheating was done first. The model parameters were extracted and optimized in the normal way for a packaged 1.6 x 5 (4m)2, junction isolated, NPN transistor obtained from Analog Devices [Get78]. All measurements were made using a temperaturecontrolled oven to supply a constant ambient temperature. The data for the characteristic curves shown in Figure 56 were measured from a 03 V sweep of the collectoremitter voltage with the base voltage held constant at 0.8, 0.85, and 0.9 V. Each point is the average of 16 measurements made after holding the bias for 20 seconds so the temperature could stabilize. The same delayandmeasure technique was used to record the Gummel plots used in the extraction. For the Gummel plots the baseemitter voltage was swept from
0.51.1 V with the collectorbase voltage held constant at 0.5 V to avoid quasisaturation. Note that it would not be uncommon to use a VCB of 1 V, doubling the power dissipation seen here.
After the device Gummel plots and characteristic plots were measured at several temperatures, the initial parameter set was extracted using the curvefitting techniques described by Getreu [Get78]. The exception to this was the base and emitter resistances were extracted using the method described by Ning and Tang [Nin84]. This is an adequate
4.0 1 W T
1.0 2.0 CollectorEmitter Voltage (V)
Figure 56 Characteristic curves resulting from a normal
model parameter extraction.
2.0 &4
0
0 U..
method for getting values for these parasitic parameters when ac methods are unavailable. The extraction was for only the parameters needed for simulation of the forwardactive and saturation regions; the reverse modes of operation were ignored because of their infrequent use in circuit design and the obvious extension from the forward modes examined here.
Next the model parameters were tuned using nonlinear optimization, the extracted data, and the output of SPICE simulations of the measurement using the parameters. Most of the parameters required only small corrections with the exception of the Early Voltage. Results of the optimization are shown in Figure 56. This simulation result is a poor fit to the measured data especially at high currents where the effects of selfheating can be seen in the increased slope of the IC versus VCE curve.
One way to correct for such selfheating errors is to measure the thermal resistance and use it with the self heating model in SPICE. The extracted effective thermal resistance using (5.9) was 800 K/W. To complete the selfheating model, the SPICE thermal parameters XTI and XTB were also extracted using the measured data over many temperatures. The resulting fits to the data are shown in Figure 57 and Figure 58. When selfheating is added to the simulation using the same model parameters, as shown in Figure 59, the conductance modeling is improved but the accuracy of the current modeling is not. It is not enough to simply include selfheating after an isothermal extraction.
The model parameters were optimized once more, this time including selfheating in the model. The Early voltage was not optimized; the value computed from (5.9) during the thermal resistance extraction was used instead. The results are shown in Figure 510. The fit to the data is much improved with good agreement in both slope and magnitude of the collector current.
The model parameter optimization including selfheating only affected three parameters significantly: IKI.F VA, and RB. The value of IKF
1012
1014 .2
cd 1015
1016
20 30 40 50 60 70 Temperature (ï¿½C)
Figure 57 Resulting model fit for extraction of XTI from the
IS versus temperature data.
in this optimization was 40% lower than that found in the isothermal extraction. Selfheating increases the collector current in the isothermal extraction, so the electrical value of kneecurrent is overestimated. A 50% increase in the electrical value VA results from taking into account the increased slope of the characteristic curves due to selfheating. Finally, a 25% decrease in the base resistance is likely caused by temperature dependence of the current gain. Although the change in base resistance
180
170 M 160
150
140 L20
30 40 50 60 70
Temperature ('C)
Figure 58 Resulting model fit for extraction of XTB from the
PF versus temperature data.
81
when selfheating is included, it is likely that RB is best extracted using highfrequency ac methods [Get78].
4.0
3.0 + + +
+
,, +
2.0  Si
0
0
1.0
001.0 2.0 CollectorEmitter Voltage (V)
Figure 59
Characteristic curves resulting from the same model parameters with selfheating added to the simulation.
4.0
3.0
+ Measured Data
Simulated
2.0
1.0
0.0
0.0 1.0 2.0 3.0 CollectorEmitter Voltage (V) Figure 610 Characteristic curves resulting from including
selfheating in the extraction process.
CHAPTER 6
SELFHEATING FOR SOISPICE
SOISPICE is a simulation tool developed at the University of Florida to do physicsbased modeling of silicononinsulator metaloxidesilicon fieldeffect transistors, SOI MOSFETs, for circuit simulation [Fos96, Suh95, Yeh95]. The basic structure for these devices is shown in Figure 61. There are two variations of this device, differing in the thickness of the body layer. In one case the body layer is thick enough so that in normal saturation region operation the depletion region does not reach the bottom of the body, and in the other case the body is thin enough for the depletion region to reach the bottom. These structures are called either partially depleted, PD, or fully depleted, FD, SOI MOSFETs.
Source Gate Drain
Back Gate / Substrate
FZ Silicon ] Silicon Dioxide D Polysilicon
Figure 61 Basic structure of a SOIMOSFET with SiO2
insulating layer.
The SOI MOSFET device is surrounded on all sides with an insulating layer, usually SiO2, that forms a barrier to heat flow. So devices tend to get much hotter than their bulksilicon counterparts, because the heat is trapped in the small silicon island the device is built in [Goo95]. The dominant temperaturedependent effect on MOSFET current is the carrier mobility, which is a much weaker function of temperature than the dominant effect (carrier concentration) on BJT current. For comparable thermal impedances and power dissipations, selfheating can be ignored in a bulk MOSFET when it cannot in a BJT. Because of the insulation layer surrounding a SOI MOSFET, the thermal impedance can be one or two orders of magnitude greater than that of a comparable geometry bulk device. This increase in thermal impedance causes generation of much greater temperatures within the SOI MOSFET, and even the weak temperature dependencies of the transistor become significant.
SelfHeating Implementation
The following shows how the selfheating model was implemented into SOISPICE, and then some simple circuit simulation examples are shown. The implementation has many similarities to the BJT approach, but some major differences exist. Some of the differences result from the organization of SOISPICE. Other differences were requested by SOISPICE users, who preferred a singlepole thermal impedance, implemented internally, without additional inputfile circuit elements, as
was done in the BJT implementation. The organization of the SOI MOSFET model allows all of the current and voltage derivatives to be easily computed numerically. This numerical approach was a necessity, because few of the physical model equations have closedform solutions for their derivatives. The following discussion will focus on these features that were unique to the SOISPICE selfheating implementation. The details of the SOI MOSFET electrical modeling are still under development and are beyond the scope of this work, so they will be ignored when possible. Detailed FORTRAN code modifications to SOISPICE for this selfheating model can be found in Appendix C. The SOI MOSFET Model
For reference, the schematic for the SOI MOSFET model is shown in Figure 62. The model has many current sources, but most are reactive and dissipate no real power. The dominant sources of real power are the channel current, ICH, and the parasitic BJT current, IBjp The other real current sources, IRGt and IGi are orders of magnitude smaller and do not contribute significantly to the power dissipation. Other sources of power dissipation include the parasitic resistances, specifically the drain and source resistances. These may not always be negligible in computing the total power dissipation, but a problem arises with the thermal impedance modeling if they are included. The thermal impedance modeling assumes the power is generated in the channel region under the gate, and not in the large adjacent source and drain regions. However, since this
implementation uses a simple, less accurate, singlepole model, including the source and drain resistances in the power computation adds more to the total accuracy than the thermal impedance modeling deficiency takes away. Therefore, these resistive power sources were included. Specifying SelfHeating
The thermal subcircuit for the SOI MOFET is shown in Figure 63. The ground connection, TA, represents the ambient temperature of the circuit defined globally in the input file. No external connections are
ICH
4B
Figure 62 Schematic for the SOISPICE model [Fos96].
allowed to the thermal circuit in this implementation so the device description line becomes:
Zxxxx nd ngf ns ngb [nb] [BJT] model RTH= CTH= L= W= ...
where RTH and CTH are two new deviceline parameters that describe the thermal subcircuit. If the parameter RTH is not specified on the device line, the thermal subcircuit is not generated, and the standard isothermal model is used. Since there are no external connections to the thermal subcircuit and the ground node is already defined, only one new node is needed to define the temperature in the selfheating model, so only one new row and column are added to the device matrix.
Computing the Derivatives
SOISPICE computes all of the current and charge derivatives with respect to node voltages numerically. The currents are computed from the model equations once at the bias point of the current iteration. For each node voltage the model is recomputed with a perturbation, and a finitedifference approximation to the derivative with respect the node voltage is computed. For selfheating this process is extended to include another Source
Gate A I
T Power RTH CH
Drain
Figure 63 The thermal subcircuit added to the
SOIMOSFET model for simulating selfheating.
perturbation computation for the temperature, and the power is added to the list of currents. For this process to have the desired results, the model equations must include a power computation and the equations for the model parameter temperature dependence. Computing the Power Dissipation
The equation used for the power dissipation is
P = IcH+ IBJTIVds + IICH + IBJTI2(RD + Rs + RLDD + RLDS), (6.1) where ICH is the channel current, IBJT is the parasitic BJT collector current, RD and RS are the drain and source resistances, and RLDD and RLDS are the low doped drain and source region resistances. The other modeled currents that are not included in this power dissipation computation, the thermal generation and impactionization currents, flow in regions away from the channel and are an order of magnitude smaller than the two included currents. Local Temperature Update of Parameters
The second task of including the temperature dependence of the model parameters is achieved using a new subroutine. This subroutine contains all of the SOI MOSFET temperature equations found in the global temperature update routines. These equations have been optimized to reduce computation time. The parameters affected are shown in Table 61 with a short description of each one. These dependencies were implemented for the SOISPICE global temperature model by Glen Workman [Wor96]. Equations for both the partially depleted and fully depleted devices are included in the subroutine. A flag is used to select between
Table 61
List of model parameters with temperature dependence.
Parameter Brief Description
Name
XNIN Intrinsic carrier density EGFET Silicon energy gap VTT Local Thermal Voltage XMULDSJ LDD carrier mobility XMUBH Minority carrier mobility in highdoped body UO Zero field electron mobility VSAT Carrier saturated drift velocity WKF Frontgate work function difference WKB Backgate work function difference VFBF Frontgate flatband voltage VFBB Backgate flatband voltage JRO Bodysource junction recombination current coefficient ALPHA Impactionization parameter cxo TAUG Carrier lifetime in lightly doped region VBI Source to bulk builtin potential PHIB Twice Fermi potential of body RLDD LDD sheet resistivity BETA
FSAT fsat
equations that the PD and FD devices do not have in common. The result is that only one routine must be modified if the temperature equations are modified in the future. As was the case in the BJT selfheating implementation, the thermal voltage must be treated with special care. This parameter is defined globally for all devices in SPICE, and a local
value (shown as VTT in Table 61) must be defined for the model equations.
Loading the Jacobian
The loading of the selfheating elements into the Jacobian matrix is done methodically. The current derivatives with respect to temperature are loaded in the new temperature column at the rows corresponding to the current branch nodes with the sign indicating direction of flow. The power derivatives with respect to voltage are loaded in the temperature row at the columns corresponding to the two nodes that comprise each voltage. The thermal conductance is loaded at the temperature diagonal element. The terms associated with the thermal capacitance are not loaded at this point, but instead are handled during preprocessing. The Jacobian matrix load for the ac analysis is implemented in nearly the same way, using complex variables for the reactive elements instead of the currentconductance pairs of the transient analysis. The simplified flowchart of how the thermal model implementation works is shown in Figure 64.
SOISPICE already has the capability to deal with linear capacitors. It is a burden on the programer to implement all of the state variables and computations to provide for the thermal capacitance inside the SOI MOSFET routines. By developing a way to trick SOISPICE into believing that the thermal capacitance is the same as any other capacitance, SOISPICE processes the thermal capacitance with all of the other circuit
linear capacitor elements. This is done by adding a feature to the SOISPICE FIND subroutine. The last argument of the subroutine call, IFORCE, is used either to return the location of a device in memory or create space in memory and store a device there. The latter option checks whether the device has an unique name and returns an error if it does not. A third option was added that stores the device without checking the name uniqueness. Using this option the thermal capacitors are all added to the
READI Decode device line including RTH and CTH
ICheck if thermal model is needed.
If so, generate thermal subcircuit.
FIND Add CTH to the list I of circuit capacitors.
STMPUPD Update model parameters for ambient
TMPUPDtemperature.
NF FET StrtNewton loop..
Coimute derivatives for the NFJaco ian including temperature.
Evaluate model at local
_ _ i I;oeotemperature.
SOITMP
New routine to update moelprameters for
local temperature
[Done
Figure 64 Flowchart of SIOSPICE execution showing
processing done for selfheating.
bottom of the normal capacitor list with the proper connections to their thermal subcircuits, all having the same name to avoid possible name collisions with the regular circuit capacitors. SOISPICE then processes the thermal capacitors like any other linear capacitor in the circuit without adding more code to the SOI MOSFET model routines.
This approach also improves efficiency. Since any computations added to the model routines are executed twice per iteration to compute the partial derivatives, excluding the computations associated with thermal capacitance from the model routines reduces computational overhead.
Convergence Issues
Because of floatingbody effects and other complex SOI MOSFET effects, convergence control and initial conditions must be set with care even without selfheating to achieve convergence of some circuits. Selfheating tends to degrade the robustness of convergence, but this is expected any time the dimension of the Jacobian is increased as it is for selfheating. The difficulty of estimating the extent of the degradation is exacerbated by the convergence behavior of the isothermal model, which was being refined concurrently with the selfheating implementation. Either the technique of skipping the first Newton iteration as used in the BJT implementation or a traditional temperature stepsize limiting algorithm could be incorporated to enhance convergence.
Simulation Results
Three example circuits were simulated to illustrate some of the effects of selfheating. The first is a dc simulation to generate the characteristic curves. The commonsource configuration is used with the sourcedrain voltage swept from 0 to 3 V for three different gate voltages. The results for the floating body case are shown in Figure 65.
5.0 .
Selfheating
4.0   No Selfheating
.3.0
2.0
1.0
0.0 _
0.0
Figure 65
1.0 2.0 DrainSource Voltage (V)
DCsweep simulation for an nchannel, floatingbody, PD, SOI MOSFET with 0.2 pm length and 10 pm width. VDS was swept from 0 to 3 V for VG of 1, 1.5 and 2 V. The thermal resistance was 5 kW/K
The second is a stepresponse test of a single nchannel, PD, SOI MOSFET. Again, the commonsource configuration is used with the gate biased for the saturation region. The drain voltage is then pulsed. The resulting drain current response is shown in Figure 66. Unlike the BJT behavior, the current decreases from the effects of selfheating. The dominant temperaturedependent carrier mobility decreases with temperature and the drain current along with it. Other than the obvious
4.5 1
I
Selfheating
 No Selfheating
4.0
0J
3.
3.0
0.0
Figure 66
0.1 0.2 0.3 0.4
Time (s)
Transient current response to a drain voltage step for a commonsource, nchannel, 0.2xlOum, PD, SOI MOSFET. Vg=2V Vds=23V Rth=5kW/K

Full Text 
PAGE 1
MODELING OF TRANSISTOR SELFHEATING FOR CIRCUIT SIMULATION By DAVID T. ZWEIDINGER A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 1997
PAGE 2
ACKNOWLEDGMENTS I would like to thank Dr. Robert Fox for his patience, the Semiconductor Research Corporation for the financial support, the many people of Harris Semiconductor for their assistance and encouragement, Jon Brodsky for his long hours simulating heatflow, and finally my parents for doing it all before me and reminding me why it was worth the work. ii
PAGE 3
TABLE OF CONTENTS pa ge ACKNOWLEDGMENTS ii ABSTRACT v CHAPTERS 1 INTRODUCTION 1 2 BACKGROUND 4 The SPICE GummelPoon Model 4 Thermal Impedance Modeling 6 SelfHeating Models 10 3 MODELING FOR CIRCUIT SIMULATION 15 An Overview of SPICE 15 Formulation of the SelfHeating Model 17 Implicit Temperature Approach 18 Thermal Impedance Modeling 19 The DC Solution 20 AC Modifications 27 Advantages and Limitations 28 Explicit Temperature Node Approach 29 The DC Solution 29 Implementation Options 32 SPICE Code Modification ' ' ' ' 34 Advantages and Limitations 39 Convergence 40 4 SELFHEATING IN CIRCUITS 44 SelfHeating Effects 44 Simulation Examples 4g Convergence Comparison 57 iii
PAGE 4
iv 5 MEASUREMENT OF MODEL PARAMETERS 60 Measurement of the Thermal Impedance 60 Temperature Coefficient Measurement 62 Thermal Impedance Measurement 67 Thermal Impedance Model and SPICE 72 Selfheating and Model Parameter Extraction 75 6 SELFHEATING FOR SOISPICE 83 SelfHeating Implementation 85 The SOI MOSFET Model 85 Specifying SelfHeating 86 Computing the Derivatives 87 Computing the Power Dissipation 88 Local Temperature Update of Parameters 88 Loading the Jacobian 90 Convergence Issues 92 Initial Simulation Results 93 Conclusion 97 7 CONCLUSION 98 APPENDLX A 101 APPENDIX B 115 APPENDLX C 127 APPENDIX D 140 BIBLIOGRAPHY 148 BIOGRAPHICAL SKETCH 153
PAGE 5
Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy MODELING OF TRANSISTOR SELFHEATING FOR CIRCUIT SIMULATION By David T. Zweidinger August 1997 Chairman: Robert M. Fox Major Department: Electrical and Computer Engineering Semiconductor devices consume power during operation. This power is dissipated mostly in the form of heat that increases the temperature of the device. This effect is known as selfheating. For bipolar transistors, BJTs, and silicononinsulator metaloxidesemiconductor fieldeffect transistors, SOI MOSFETs, selfheating can significantly change device performance. Circuit designers need an easy way to quantify the performance changes from selfheating, within the compact models used in today's circuit simulators. A historical review of other attempts at modeling selfheating in circuit simulators is shown. Two approaches to including selfheating effects for bipolar transistors in the popular circuit simulator SPICE are presented. One approach uses an implicit temperature and transcendental equations to correct for selfheating in the device model. The other v
PAGE 6
approach uses an explicit temperature and an electrical subcircuit analogy for power generation and thermal dissipation to model selfheating. Techniques for assessing selfheating effects in hand calculations are shown, and example simulations with selfheating in common dc, ac, and transient circuits are shown. Reduced amplifier gain and thermal mismatching are common selfheating effects. A new method for extracting the thermal impedance of bipolar transistors is derived, and results are shown for transistors of several different geometries. The effects of selfheating on the extraction of the GummelPoon model parameters are explored using longtime measurements and the modified SPICE in the optimization process. The implementation of selfheating in the physicsbased SOISPICE model for SOI MOSFETs is shown. Common circuit examples are simulated showing the effects of selfheating. For analog circuits, the negative temperature coefficient of drain current can cause problems for biasing and amplifier gains. For digital circuits, selfheating is a function of frequency and is not significant. vi
PAGE 7
CHAPTER 1 INTRODUCTION In present circuit simulators, temperature is primarily modeled globally. Only a single temperature is used for all of the circuit elements, and the circuit is simulated isothermally at that temperature. This approach ignores selfheating, the local increase in temperature from the power dissipation of a device. The operating point of a device defines its power generation, and the geometries and materials affect how the power is dissipated as heat. The isothermal simplification is inaccurate considering that most circuits have many elements of different geometries, operating under different bias, and dissipating different amounts of power. The result is a performance change of the circuit because of all of elements operating at different temperatures. These performance changes from selfheating are significant for many circuits. Semiconductor devices are especially sensitive to their operating temperature. One device that is particularly sensitive is the bipolar junction transistor (BJT). In BJTs the dominant sensitivity arises from the exponential dependence of the carrier concentration on temperature. Only a few degrees Kelvin of temperature change will cause significant changes in the carrier concentration. The BJT terminal currents are a function of the carrier concentration, and hence are greatly affected by 1
PAGE 8
2 small temperature changes. Some temperature dependent mechanisms that are weaker than the carrier concentration are the carrier mobility and recombination lifetimes which become significant when large temperature changes occur on the order of tens of degrees Kelvin. These two mechanisms are important to silicononinsulator metaloxidesilicon fieldeffect transistors, SOI MOSFETs. Selfheating can cause failure of heterojunction bipolar transistors (HBT) where the large negative temperature coefficients cause currents to drop to zero in relatively lowpower operation [Gao89]. Including these effects in the compact BJT and SOI MOSFET models used in circuit simulation is important for improving accuracy in some circuits. The importance of selfheating has become more evident with changes in technology. The continued shrinking of geometries, and the resulting increase in current densities and restriction of heat flow, tend to increase the thermal impedance and raise device operating temperatures. The thermal impedance is the temperature increase for a given power dissipation. The trend toward the use of silicon dioxide, Si0 2 , for isolation makes selfheating more important, because the decreased thermal conductivity of Si0 2 compared to silicon creates a barrier to heat flow and greatly increases the thermal impedance of the device. Finite elementbased simulators are available that can simulate a BJT under dc and ac conditions with selfheating [Lia94], but these are extremely slow for analysis and design of circuits with multiple devices.
PAGE 9
3 What circuit designers need is a compact modeling simulator that includes selfheating. SPICE, with its GummelPoon model, is a popular circuit simulator. The temperature dependencies of all the parameters are implemented in the code for simulation of circuits at different global temperatures. The goal is to implement selfheating in the model by incorporating the temperature dependencies and the device thermal impedance and power dissipation to allow each device to be simulated at its own local temperature. The following chapters are organized to give a understanding of the motivations for this work and the solutions reached for incorporating selfheating into circuit simulation. Chapter Two a historical view of thermal impedance modeling and selfheating simulation. Chapter Three describes two ways to incorporate selfheating into SPICE. Chapter Four focuses on circuit analysis in the presence of selfheating and shows some example simulations. Chapter Five shows a new extraction method for measuring thermal parameters of BJTs, in particular the thermal impedance. The effects of selfheating on model parameter extraction are also examined. Chapter Six shows the implementation of a selfheating model for SOI MOSFETs, with some simulation examples included. Finally, Chapter Seven summarizes the information and insights presented.
PAGE 10
CHAPTER 2 BACKGROUND This chapter presents an overview of previous work in modeling selfheating. First, the BJT device model equations used as the framework for selfheating simulation are reviewed. Following this is an investigation of thermal impedance modeling. Finally, an examination of previous attempts to model selfheating in BJTs will show the need for improvement. The SPICE GummelPoon Model The dominant model for use in simulating bipolar transistors in circuits was introduced in 1970 [Gum70]. This integral chargecontrol model was incorporated into the circuit simulator SPICE, developed at the University of California at Berkeley. Most circuit simulators in use today have a core based on SPICE and its GummelPoon BJT model. Therefore, SPICE was the obvious platform for the study of selfheating in BJTs and circuits. The communication of selfheating modeling is helped by the common use of SPICE based simulators in industry. What follows is a short overview of the SPICE GummelPoon model that will form the basis for most of the discussion in the chapters to come. The following is the GummelPoon model for an npn BJT. The base current (I B ) is composed of four current elements, 4
PAGE 11
5 l B " I BF + I BE + I BR + I BC (2.1) where Igp is the forward diffusion current that results from backinjected holes across the baseemitter junction, given by N F V T 1 (2.2) where Ig is the saturation current, (3pis the forward dc current gain, Np is the forward current emission coefficient, and V T is the thermal voltage. The nonideal baseemitter current results from recombination in the quasineutral base, and is given by: ( Vbe \ N E V T 1 (2.3) where Ig E is the baseemitter leakage saturation current and N E is the leakage emission coefficient. Similarly, the dual for the reversebias case exists. I BR is the reverse diffusion current given by: ( V BC N R V T BR (3 R 1 (2.4) where (3 R is the reverse dc current gain and N R is the reverse current emission coefficient. Finally the nonideal basecollector current is ! BC ~ J SC nc N C V T 1 (2.5)
PAGE 12
6 where Igc is the basecollector leakage saturation current and Nq is the leakage emission coefficient. The collector current is given by T Â— c Q B N F V T e N R V T _I BR _I BC (2.6) where Qb is the base charge factor that accounts for basewidth modulation and highinjection conditions. The base charge factor is Qb = T^V 17 ^ (2.7) where Q. = Xbc_Xbe vI V AF V AR. (2.8) and f VfiE A O 5Â§ KF N F V T 1 ( Vbc A + L sc KR N R V T 1 (2.9) The parameters and are the forward and reverse early voltages, and Ikf and Irr are the forward and reverse knee currents or corners for beta high current rolloff. Thermal Impedance Modeling As a device heats from its power dissipation, its operating point shifts causing the power dissipation to change. This circular mechanism of heating is called thermal feedback. There are several thermal feedback
PAGE 13
7 mechanisms in integrated circuits, which can usually be considered separately [Mei77, Fox91c]. One mechanism is the rise in the overall chip temperature due to the total power dissipated on the chip. This temperature rise is controlled by the chiptopackage and packagetoambient thermal impedances. This global heating operates over a long time scale (milliseconds to minutes), and couples all of the devices on a chip. Global heating can generally be reduced by careful packaging and heatsinking, and is accounted for in circuit simulation by simulating at multiple temperatures. Global heating is important to recognize in parameter extraction, where long thermal settling times cause the isothermal assumption to become inaccurate. For largearea devices, or those dissipating large power, there can be direct coupling between the heat dissipated in one device and the temperature of other devices. This mechanism is strongly affected by the circuit layout. For example, thermal coupling in an operational amplifier from the output stage back to the input can profoundly affect the circuit's gain characteristics, although careful layout exploiting symmetry can mitigate these effects [Fuk76]. Direct heating of a transistor by its own power dissipation cannot be eliminated through packaging, scaling, or layout. This mechanism can cause substantial errors in modeling even without high power dissipation. As shown by Joy and Schlig, for modest power dissipation, the temperature rise is mostly confined within the transistor itself [Joy70].
PAGE 14
8 Thus this effect can be simulated only by allowing each transistor to be modeled at its own temperature, controlled by its thermal spreading impedance and power dissipation. The thermal spreading impedance, which controls the temperature rise, can be predicted from the transistor's geometry [Lee92, Bro93, and Zwe95b]. The thermal impedance model derived by Fox and Lee [Fox91b] is the frequencydomain equivalent of the timedomain derivation shown by Joy and Schlig [Joy70], and is based on the same assumptions. All of the heat is assumed to be generated uniformly in the rectangular volume of the collector spacecharge region (SCR). The analysis assumes an adiabatic plane at the surface of the transistor created by an image of the SCR heatsource placed above the surface. This approach assumes there is little heat lost through the top of the BJT and the heat is removed predominately by the conduction of the body of the semiconductor. For isolation technologies where Si0 2 is used between devices, this may not be a good assumption [Dav92, Bro97a]. In the time domain, the response to a point impulse of heat, measured a distance r away in a uniform medium, is where K is the thermal conductivity and k is the thermal diffusivity. The Laplace transform of (2.10) gives the thermal impedance For s = 0 , this equation gives the thermal resistance R TH for the point source. (2.10) (2.11)
PAGE 15
An expression for RpH a t any point r' = (x', y', z') in the semiconductor due to the heat generated in the collector SCR can be derived by integrating Rj^ over the collector SCR and its image source. For a rectangular vertical BJT the integral is R^W.L.D.H.O = ^jJ^^ + J^fififf] ,2.12) where r = J(x x') 2 + (y y') 2 + (z z') 2 , W and L are the width and length of the emitter, D is the collectorbase junction depth, and H is the SCR width. The result in (2.12) can be numerically approximated by [Fox91a, Fox91b]: R TH = =L (2.13) 4nKVWL Â• f j Â• f 2 where f^d.h) = (0.058d + 0.14)h + 0.34d + 0.28 , (2.14) f 2 (a) = 0.98 + 0.043a 6.9 Â• 10" 4 a 2 + 3.9 Â• 10" 6 a 3 , (2.15) d = ^L, (2.16) W a = ,a>l, (2.17) Â„ = ^(VcB^)AqN Ep! ) N EPI is the epitaxial doping density in atoms/cm 3 , V CB is the collectorbase junction voltage and <> bl is the collectorjunction builtin potential. This expression was found experimentally to approximate R TH well for a wide
PAGE 16
10 variety of rectangular emitter geometries [Fox91a]. Based on a comparison of (2.11) and (2.13), the same Rth could be computed by replacing the collector SCR with a point heat source placed a distance r eff below the emitter, where r e ^can be denned as r <Â« = 2FTri^ = 2 ^f 'f 2(219) This r e ftcan be used in (2.11) to simply estimate for Z 
PAGE 17
11 device flows as a current through an impedance, establishing a potential equivalent to the change in temperature. Only a single RC pole is shown, Time (s) Figure 21 Normalized temperature response to a 1 W power step for a 7xl0[im npn BJT, using the model in [Joy70] and the inverse Fourier transform of the frequency response of (2.20) [Fox93b]. Figure 22 Example of thermal subcircuit used to model selfheating.
PAGE 18
but electrical networks representing multiple poles could be constructed to better model the distributed nature of the thermal spreading impedance [Mul64]. An early attempt at computer simulation of selfheating effects was done by Latif and Bryant [Lat81]. In this work a simple, single timeconstant, thermal subcircuit was implemented into the EbersMoll BJT model [Ebe54]. They used a linear network solver (WATAND) to simulate dc and transient effects of second breakdown in power BJTs. Later, this work was extended to include the more accurate GummelPoon model in the same solver for a study of selfheating effects on current mirrors [Mun91]. These two works failed to include the effects of selfheating in the ac analysis. A singlepole thermal subcircuit was used to model temperature in SPICE by Vogelsong and Brzezinski [Vog89]. Results from a few circuit simulations were presented. Again, only dc and transient analysis were implemented, ignoring the important ac analysis. Their approach to computing the power dissipation included heat sources other than the basecollector junction which all have different thermal impedances that were not modeled. The algorithms they used to implement selfheating are vague. In all of the above implementations, values used for the thermal impedance are empirical. The impedance value used is the one that makes the simulation fit the data best. This approach can neglect the coupling of
PAGE 19
13 selfheating effects to other effects in the device behavior. Most of the implementations used only a singlepole network which poorly models the distributed nature of the thermal impedance. Finally, all of the simulators above require a voltage source in the thermal network to represent the ambient temperature. Repeating a simulation at a different ambient temperature would require modification of all of the thermal networks in the circuit. The underlying physics of the heat flow either within the transistor or between transistors is often neglected in the above works. Heatflow analysis shows that except for rather highpower transistors, temperature gradients are mostly confined within the transistor itself [Joy70]. Usually adjacent devices are thermally coupled through the devicetochiptopackage spreading impedances, that have very long time constants, rather than devicetodevice impedances [Lee92]. The heat coupling effects can usually be ignored. Where the heating of adjacent devices cannot be ignored is in multifingered emitter BJTs and HBTs. The close proximity of each finger causes a strong thermal coupling. For example, for a three finger BJT driven with a constant current, the middle finger is heated by the two outside fingers causing its current to increase and the others to decrease. This effect can degrade the performance of multifingered emitter devices. The improved implementation shown in the next chapter includes dc, transient, and ac analyses. It allows flexibility in formation of the
PAGE 20
14 thermal impedance network, so that the self impedance and impedance to adjacent devices or emitter fingers can be modeled accurately.
PAGE 21
CHAPTER 3 MODELING FOR CIRCUIT SIMULATION Because of the deficiencies of previous attempts to include selfheating in circuit simulation, a more complete selfheating implementation was needed. The first priority was to create a simulator that could model the ac effects of selfheating. The frequency domain had been mostly ignored for analog circuits. The first implementation filled this void using an implicit local temperature approach. In the implicit temperature approach the BJT model is reduced to a twoport yparameter equivalent model. The yparameters are then modified by selfheating, and the modified yparameters are converted back to the original model form. Unfortunately, this implementation made transient analysis very difficult, so another implementation using an explicit temperature and thermal subcircuit was done. A thermal subcircuit represents the power dissipation, heat flow, and temperature rise. This requires a thermal impedance network with some inherent limitations, but does allow transient analysis, unlike the implicit node approach. Both of these approaches are described in this chapter. An Overview of SPICE To understand the implementations, it is important to understand first the organization and function of SPICE. The circuit simulation 15
PAGE 22
16 program can be broken down into four pieces: preprocessing, dc and transient solution, ac analysis, and postprocessing. The preprocessing section is responsible for reading the circuit description file and loading the information into the data structure. This processing includes building a instance matrix of all the circuit elements to preserve the connectivity information, and associating any parameter values and models that go with each instance. The data structure for the Jacobian matrix, voltage and current vectors, and state variables is then built. Any processing of model parameters is done, and the global temperature dependences are updated for the selected temperature of the simulation. The dc and transient section of SPICE loads the Jacobian matrix with the circuit conductances and branch currents, and solves the resulting linear equations for the node voltages. For nonlinear elements an iterative NewtonRaphson method is employed. The device models are used in routines to compute the next guess for the conductances, currents, and voltages associated with each device. During transient analysis the energy of storage elements is integrated for each time step, and reduced to an equivalent conductance and current for the matrix load. The ac analysis section uses the real part of the Jacobian from the solution in the dc analysis, and computes the complex part of the matrix for reactive elements at a given frequency. Perturbation of a node voltage
PAGE 23
or branch current gives the smallsignal response at the linearized operating point. Finally, the postprocessing routines gather the simulation results and prints or plots the data. The results can include device operating point information, sweep data from ac, dc, and transient analyses, and simulation statistics. Formulation of the SelfHeating Model The derivation of the both modeling approaches shown here starts the same way, with the expression for the smallsignal collector current. For the current into the collector including selfheating 31 iÂ„ = C 5V BE Tiixed ^ T.fixed v cb + ap C P (3.1) where the power is P = Vce + i c V CE + I B v be + i b V BE< 3 2 ) Expanding the power term in (3.1) and reducing the partial derivatives to conductances results in <* l C 3T J c = Â«oE v ce + SmE v be + Â«uE v cb + 5f ' 3pP (3.3) where the E subscript denotes electricalonly. Substitution of the thermal resistance gives l c = S 0 E v ce + gmE v be + guE v cb + af R THP( 3 4)
PAGE 24
18 The last partial is substituted using the definition of the fractional temperature coefficient given by dl m D m = (l/TJJ? (3.5) where m is c or b for the collector or base current. The resulting substitution gives i C = SoE v ce + SmE v be + ^E v cb + ^(^THP Â• < 3 ' 6 > Equation (3.6) marks the point of divergence of the two implementation approaches. Solving for i c gives a ratio with temperature represented implicitly as done by Miiller [Mul64]. This is the basis of the implicit temperature approach [Zwe92]. If the last term in (3.6) is represented using temperature for the thermal resistance and power product, a new matrix node is needed. This is the basis for the explicit temperature approach. Implicit Temperature Approach This is the simplest modification to make to the SPICE source code. The only changes needed are the addition of the new model parameters and modification of the device model routine. There are no new model matrix elements or state variables that would require changes to the device matrix. Detailed FORTRAN code modifications to SPICE2G.6 for this approach can be found in Appendix A.
PAGE 25
19 Thermal Impedance Modeling In this approach the thermal impedance is computed for each iteration of the device solution. New model parameters to describe the geometry of the BJT are added to implement the thermal impedance equations discussed in the last chapter. These parameters are: the length and width of the emitter, the depth of the collectorbase metallurgical junction, and the collector doping concentration. They are added to the list of values available on the BJT model line. From these four parameters and the collectorbase bias voltage the effective thermal resistance is computed from equations (2.13) through (2.18) for each iteration in the dc solution. For the ac analysis, the effective heat source radius, r e ^, is computed from the thermal resistance computed in the dc solution or given in the model line. Then (2.20) is used to compute the thermal impedance at each frequency. The thermal impedance can also be computed from the thermal resistance and capacitance by addition of the thermal capacitance to the model line. This allows a simple singlepole model to describe the thermal impedance. A hierarchy was formed to manage these three models. The first option is the lowest level where Rr H and C TH (thermal capacitance) must be supplied and are used in the dc and ac analyses. The next option is used when only R TH is supplied. In this case, the given R TH is used in the dc solution, and Z TH is computed from r eff for each frequency in the ac
PAGE 26
20 analysis. The last option uses the geometry information and the predictive model to compute Rp^ for the dc solution and Z TH for the ac analysis. The DC Solution The following is a brief description of the changes made to the routine that computes the dc solution for the BJT. Including the temperature dependence A first step to all approaches to modeling selfheating is to copy all temperature dependences to the inside of the NewtonRaphson loop. Normally in SPICE the model parameters are updated for temperature once before each simulation is begun, allowing circuits to be sequentially simulated at different global temperatures. By moving the parameter dependences inside the NewtonRaphson loop, the temperaturedependent parameters for each transistor are updated for its local temperature in each iteration. The dependentparameter temperature functions are described below. The thermal voltage is implicitly temperature dependent: V T = kT/ q where k is Boltzman's constant and q is the electron charge. The temperature dependence of the dc current gain parameters stem from the dependence of the junction barrier height on temperature, and are given as (3.7) and
PAGE 27
21 T \XTB Pr = Pro[^ (3.8) where T 0 is the original model parameter temperature and XTB is the 3 temperature factor. The following current parameters dependences result mostly from the strong temperature dependence of the carrier concentration: I s = I so e' a '"Â° , (3.9) factln IsEO n f i se = Â— ^br e . (3.10) (T/T 0 ) A1M and factln ! SC = ^TBe . (3.11) hco N c (T/T 0 ) The exponential factor in the above equations is E G E G ^ T , (T factln = 2+ XTI Â• In J(3.12) v T0 V T \J { where XTI is saturation current temperature factor and T 2 E G " E G0 ~ E GAP ' fTr (31 3) 1 + ^AP is the semiconductor energy gap. Values for the arguments in (3.13) are shown in Table 31 for common substrate materials. The SPICE BJT model assumes the energy gap is constant and uses a single parameter in the model line to assign a value. This has not changed for this implementation. Equation (3.13) is added to the temperature equations for the explicit node approach
PAGE 28
22 Material E G0 (eV) E GAP (eV/K) T GA p(K) bihcon 1.16 7.02xlO" 4 1108 Germanium 0.67 4.56xl0~ 4 210 Gallium Arsenide 1.52 5.41X10" 4 204 Table 31 Energy gap temperature modeling parameter values for common device materials. to fix this deficiency in SPICE. New model parameters to assign values for the coefficients were also added to the model line. Computing the temperature The temperature is not represented in the Jacobian in this approach. Instead, it is computed in each iteration from the thermal resistance and the power dissipation, T = P Â• R TH The power is computed in each iteration of the NewtonRaphson loop from the dissipation in the basecollector and baseemitter spacecharge regions, P = IqVq + Ib^b Note this is the real power dissipation, and does not include any reactive power that may be present in a transient simulation. The thermal resistance is either provided by the user or computed from the geometry using equations (2.13) to (2.18). Modifying the Jacobian In the dc analyses the hybridrc parameters are converted into twoport yparameters using Y llE = Sn + Su (314)
PAGE 29
23 Y 12E = Si Y 21E S m g^ (3.15) (3.16) (3.17) Y 22E = So + gp where the hybridrc parameters are shown in Figure 31. Next the BJT commonemitter yparameters are corrected to include thermal feedback using [Miil64] mn 1 " D m R TH P (3.18) where m or n equal 1 for the base or 2 for the collector, Y mnE is the uncorrected electrical yparameter, and D m is the fractional temperature coefficient of the base or collector current. The collector current temperature coefficient in the forwardactive region of operation is estimated by [Zwe93] base ^VV collector emitter So /As yi2 Figure 31 Conversion between standard hybridrc and twoport yparameter models.
PAGE 30
24 D C = f l 11, (3.19) where f l = XTI + qE G /kT qV BE /kTN p (3.20) For the base current temperature coefficient in same region the expression is D B = BE f 2 v BEq N E ~ kTN E XTB + 1 bf V BE<1 ( v 2 kTN F XTB i A /(I B T) (3.21) where E Gq f 2 = XTI+ kT (3.22) The thermally corrected yparameters are then converted back to the hybridn form using the following ordered equations: SH = " Y 12 &n = Y ll Â«n Â«m = Y 21 + Â«n (3.23) (3.24) (3.25) g0 = Y 22"^ (326) The hybridK parameters are then loaded into the admittance matrix. Again, the advantage here is that no new matrix elements are needed in the Jacobian to compute the selfheating effects.
PAGE 31
25 Flow control The flowchart of the dc solution routine with selfheating is shown in Figure 32. First, the voltages from the Jacobian matrix solution are loaded, and the currents are linearly predicted using the conductances, the change in node voltages from the previous iteration, and the previous iteration currents. Using these predicted values of current and the node voltages, the power is computed, followed by the temperature using Rth(If the predictive R^h model is used then the value for Rr H comes from the previous iteration.) Next, the convergence of the currents, voltages, and temperature is checked. The predicted currents and the node voltages are compared to the values of the previous iteration, and if each comparison is within its specified tolerance, the device has converged. If the convergence check succeeds then the statevariables are saved, the BJT is marked as converged and the model routine is exited. If the convergence criteria are not met then the predicted temperature is used to update the parameter temperature dependences and the currents are computed from the model equations described in Chapter 2. With the currents computed from the model, the power, Rr H , the temperature, and the fractional temperature coefficients are computed. The convergence is checked again for the new current values against the predicted values computed earlier. If the convergence check succeeds then the state variables are saved and the BJT is marked as converged. If it
PAGE 32
26 Loop begin Load voltages from matrix solution Compute current predictions 1 Compute t( from last imperature iteration r Check convergence of predictions against last iteration ~l Recompute temperature dependences Compute currents from model 0 bu u > d o O Compute temperature and impedance Solve linear equations z Load Jacobian z Transform back to hybridn I Correct for selfheating A Transform hybridK to yparameters A Save state variables i Check convergence of currents against predictions k 1 Save state variables Done Figure 32 Flowchart of the dc solution for the implicit node selfheating approach.
PAGE 33
27 fails the statevariables are saved and the matrix load begins. The BJT conductance values are transformed into yparameters using (3.14) to (3.17), corrected for selfheating using (3.18), and converted back to hybrid7i form using (3.23) to (3.26). The modified conductances are then loaded into the Jacobian. After all the BJTs in the circuit have been processed and loaded into the Jacobian, the linear equation solver is invoked to generate a new voltage solutions. SPICE continues to iterate until all the nonlinear devices converge, or a maximum iteration count is reached. AC Modifications For the ac analysis the smallsignal power is computed from (3.2). The hybrid;: model is transformed into complex yparameters as follows: Y 11E = Sn + gn + j w ( C xcb + C be + C bc) (3.27) Y 12E = gn + j w ( C xcb + C bc ) < 328 > Y 21E = Â«m "Â«n + j( x S m " o>C bc ) (3.29) Y 22E = So + Sn + j^'Cbc (3 30) where C be and C bc are the baseemitter and basecollector capacitances respectively, xgm is the excess phase term, C xcb is a baseemitter transcapacitor controlled by the basecollector voltage, and the other terms are standard hybridrc parameters shown earlier. The complex form of (3.18) where Z TH is substituted for is used to modify the yparameters to include selfheating. The fractional
PAGE 34
28 temperature coefficients used are from the dc solution. The thermal impedance is computed from and C^h given by the user or predicted from the Rth computed in the dc solution. The real parts of the yparameters are transformed back to the real parts of the hybridn model much as in the dc analysis. To transform the imaginary parts, first (3.30) is solved for C^. Then C xc b and xgm are solved from (3.28) and (3.29). Finally, C be is solved from (3.27). After transforming the modified yparameters back to the original parameter set, the Jacobian is loaded and the matrix is solved for the current frequency. Advantages and Limitations This implementation is easy to use; with only a few new model parameters the user can model selfheating. The changes to the original SPICE source code are minimal as well, making the model easy to support. The biasand frequencydependent thermal impedance model makes for smooth, accurate dc and ac simulations. The selfheating model in this implementation is only valid in the forwardactive region of operation. The most significant selfheating occurs in the forwardactive region, where the collectorbase junction has a large reversebias voltage and the high collector current create large power dissipation. The forwardactive region is also the most common region for analog applications where selfheating has the greatest effect on
PAGE 35
29 circuit behavior. It is for these reasons that this implementation has focused on this region. The greatest drawback to this implicit temperature approach is the inability to do transient simulations. SPICE is illequipped to solve the problem for this modeling approach, and this is the ultimate reason why the explicit thermal node method was implemented. Explicit Temperature Node Approach In the previous approach, the temperature was computed implicitly in each iteration using the voltage and currents of each iterate, and then modifying the existing internal conductances. In the following approach, temperature is represented explicitly by the addition of new elements to the device model, and a thermal subcircuit to model heat flow. A new row and column was added to the Jacobian along with new state variables for temperature and power dissipation. The dc, ac and transient analyses can be all implemented using this thermal node approach. Detailed FORTRAN code modifications to SPICE2G.6 for this approach can be found in Appendix B. The DC Solution The following is a brief description of the changes made to the routine that computes the dc solution for the BJT. Starting where the derivation of (3.6) left off, the new matrix elements are derived. If the substitutions of collector thermal transconductance, g TC = I C D C , and temperature,
PAGE 36
30 T = R TH P, (3.31) are made in (3.6) then 1^ becomes ic = So v ce + Sm v be + Vcb + STC T Â• (332) A similar approach used for the base current results in the expression [ b = gn v beVcb + STB T (333) where g TB = I B D B is the base thermal transconductance. Using the electrical analogy for (3.31) results in modeling the temperature as a voltage and the power as a current. To include the power in the Jacobian, it must also be represented as a linear function of voltages and conductances. Substituting (3.32) and (3.33) into the expression for power in (3.2) gives the following: P = (So v ce + gm v be + Vcb + S TC T ) V CE < 3 34 > + (Â«n v be " Vcb + S TB T ) V BE + ^ce + Vbe Using the above equations (3.31) through (3.34) the intrinsic BJT element stamp is formed: 8m + So + Sn g H "Sin "So S T C " g n s^ + g n g n g TB g m g c s n (g m + g G + gÂ„) (g TC g TB ) 1 v b i C + i b T P (3.35) gpb gpc gpb R TH + gpt where g pc = v ce (g m + g Q + g^) is the collector power term, g pb = v be (g n gjl ) is the base power term, and g pt = v be g TB + v ce g TC is the temperature power term.
PAGE 37
31 The device temperature change is computed at the new device node by a controlled current source equivalent to the device power dissipation in series with the effective thermal impedance as shown in Figure 33. The ambient or global temperature of the circuit is represented by ground, and the T A node in the figure is grounded for most cases. The power dissipated in the transistor flows through the thermal impedance producing the effective temperature change of the device. The base and collector currents are then modified by the new thermal transconductance elements and the temperature. The thermal resistance is the sum of the resistances in the thermal impedance network, shown A/Y R, 'g TB AT Srn Vbe So e Cthi IE C TH2 Figure 33 BJT internal largesignal circuit model showing new temperature node and thermal impedance ladder network.
PAGE 38
32 here divided into three poles. Normally the thermal impedance network is connected to ground at the point T A , but it can be coupled to other devices through an additional network representing the impedance between B JTs. Implementation Options Two choices affected the selection of an implementation strategy. The first choice was how to implement the thermal impedance subcircuit. The second choice was how to compute the temperature derivatives. One way to implement the thermal subcircuit is similar to the implicit node approach above and would entail coding the subcircuit inside of SPICE. This requires the BJT matrix size to increase by a row and column for each new node in the impedance network. For maximum accuracy this could require as many as five new nodes, which would almost double the size of the BJT matrix. The topology of the impedance network would also have to be chosen in advance. The form of the ladder network shown in Figure 33 or the distributed form shown in Figure 34 could be used. Once implemented in the code, whichever topology is chosen cannot be changed by the user. The advantages of the internal approach are its Rthi R TH2 Mr TH1 'TH2 y TH3 T A Figure 34 Distributed form of the thermal impedance subcircuit.
PAGE 39
33 transparency to the user and the ability to use the biasdependent dc and ac impedance model implemented in the implicit node approach. The other way to implement the thermal subcircuit is to leave it to the user to specify the thermal impedance network in the circuit description file just like any other circuit elements. With this approach the flexibility to use any topography is retained. Computing the thermal impedance is a task well suited for preprocessing using the SPICE input processing routines. The biggest advantage to this approach is that relatively few changes are needed to the BJT matrix and SPICE data structures, making the implementation much simpler. The temperature derivatives can be computed analytically or numerically. When this work began the main interest was in selfheating effects on dc and ac simulations in the forwardactive region where the effects are greatest. An analytical solution of the temperature derivatives was practical when considering only the forwardactive region, and made for an elegant implementation in the implicit node approach. However, a problem arose when a device passed through operating regions where the derivatives were incorrect. The simple problem of implementing a single region became complex when analytical expressions for other regions had to be implemented. A numerical approach to computing the temperature derivatives improves the problem of regional boundaries. The numerical derivatives are computed using
PAGE 40
34 Â— * Â— (3.36) dT TT where x is any temperature dependent variable at the temperature T and x' is the value of that variable at temperature T a small increment above T. To implement the numerical derivatives requires each variable, in this case the collector and base currents, to be computed twice, once at each temperature. This takes about the same amount of computation as the direct analytical approach because of the complexity of the analytical expressions (see equations (3.19M3.21)) compared to those of the currents. SPICE Code Modification This section details the implementation. While many different approaches have been tried over the course of this work, only the final version is covered here. In this implementation the best tradeoffs for efficiency, flexibility, and convergence have been made. The thermal impedance is described externally in the SPICE input file, and the derivatives of the collector and base currents with respect to temperature are computed numerically. Since the thermal impedance implemented externally to the BJT model, there are no new model parameters. Only one new device matrix node has been added for the temperature and four new state variables: temperature, power dissipation, and the derivatives of the collector and base currents with respect to temperature. The modifications to the SPICE data structure are relatively small, but are complicated to realize
PAGE 41
35 because of the complexity of the SPICE data structure. Fifteen subroutines were modified and one new routine added to incorporate the changes needed for selfheating, but not all of these are discussed here for the sake of brevity. For more details, see Appendix B. The format of the BJT device input line was changed to include the external thermal node. The new BJT line is as follows: Qxxxx nc nb ne [ns [nt]] model [area] [initial conditions] where the items in brackets are optional. The substrate node number (ns) and the temperature node number (nt) are both optional. The substrate node must always be given if the temperature node is used or differentiating between them when only one is specified would be impossible. An example circuit description with selfheating is shown in Figure 35 where the substrate node has been grounded. If the thermal node is not specified the node is assumed grounded and the standard isothermal model is used. Once the preprocessing routines were modified to recognize the thermal node, build the new device matrix row and column, and add the new state variables, the work of modifying the device model routine for dc and transient analyses could begin. Note that the temperature state VCE 1 0 DC 5 VBE 2 0 DC 0.8 Ql 1 2 0 0 3 NPNMODEL RTH 3 0 500 Figure 35 Sample modified BJT circuit description for SPICE with a thermal resistance of 500 K/W.
PAGE 42
36 variable stores the change in temperature above the circuit ambient, and is converted to the absolute temperature only where needed in the model. Unless otherwise stated, the temperature referred to in the following is the temperature change. In the device initialization phase, the temperature is set to zero. No power has been computed so no temperature change can exist. Next in the linear extrapolation section the equation for predicting the power is added and temperature is added to the current prediction equations: P = P + I C AV CE  + I B AV BE  (3.37) where P is the predicted power and P is the power from the previous iteration, l C = I C + (gm + 8o) AV BE(So + gm) AV BC + gTC AT ' < 3 38 > and l B = l B + Sn AV BE Â§n AV BC + S T B AT (3.39) Next the initial convergence check is changed to include the temperature. The absolute tolerance on the temperature multiplies the conversion factor of 100 K/V and the tolerance of the junction voltages. With only a unity conversion factor, using the default voltage tolerances would require the temperature to be accurate to 50 uK for convergence. This is more precision than is needed and discourages convergence. Scaling the tolerance to 5 mK provides enough accuracy and improves the robustness of convergence. The power does not need to be included in the
PAGE 43
37 check, because it is a linear function of currents and voltages and will converge when they do. Similar to the implicit node approach, the temperature update equations were moved inside the model routine. The model parameters are adjusted for local changes in temperature in each iteration. After the initial convergence check and any junction voltage stepsize limiting, the temperature dependences are inserted. A loop is formed around the temperatureupdate equations, so they can be used in computing the numerical derivatives of the currents with respect to temperature. (See the flowchart in Figure 36.) The temperature in augmented by 0.1 K, the model parameters are adjusted for temperature, and the currents are computed from the model equations. At this point the hightemperature current values are temporarily stored and the 0.1 K is subtracted back from the temperature to return to the original value. The loop returns to the temperature update and then recomputes the currents for the proper temperature. The loop is exited and the current derivatives, g TC and g TB , are computed using the approximation of (3.36). The choice of 0.1 K for the temperature change was arrived at empirically; it results in enough change in the currents so the numerical problem of subtracting two close numbers of finite precision does not occur, and it is small enough to preserve accuracy in the approximation. The power dissipation must be computed before any changes from reactive currents in the transient analysis can change the base and
PAGE 44
38 Loop begin Load voltages from matrix solution Compute current predictions Compute power prediction 1 Check convergence of predictions against last iteration Save state variables Done Start derivative loop Add delta temperature Correct parameters for temperature Subtract delta temperature Compute derivatives 1 Compute currents from model One time * Save currents Check convergence of currents against predictions Save state variables Load Jacobian Solve linear equations Figure 36 Flow chart of BJT model routine.
PAGE 45
39 collector current value. The changes in current from reactive elements in the transient analysis are not real and do not contribute to real power dissipation. Such charging currents must be dissipated in the device parasitic resistances to generate temperature changes. Since the reactive elements of the thermal impedance are denned externally, no changes are needed to the transient analysis section of the routine. SPICE includes a second convergence check after the transient section to compare the currents newly computed from the model to the predicted currents. As noted previously, it is not necessary to check the power; the temperature will not change within the model routine. The state variables are saved and the Jacobian loaded as shown in (3.34) and (3.35). The changes to the ac analysis routine are minimal. In addition to the normal matrix load, the temperature row and column are loaded with the smallsignal equivalent of the dc matrix load. Advantages and Limitations Many of the advantages and limitations to the external thermal impedance approach have already been mentioned, but is worth restating them. The advantages are many. Since the thermal impedance is implemented outside the BJT model, any topology or number of time constants can be used for the thermal equivalent circuit. No new model parameters and only a few new state variables are required; an internal thermal impedance would have many of each. The temperature
PAGE 46
40 derivatives are computed numerically, improving continuity across operating region boundaries with little or no additional computational overhead. Transient analysis is possible where it was not in the implicit node approach. The biggest limitation is that the thermal impedance is fixed. Neither bias nor temperature dependences are implemented, and the smooth frequencydomain solution of (2.20) is unworkable in the smallsignal analysis. Versions of SPICE allowing nonlinear equations in userdefined controlled sources would make it possible to overcome this limitation. Convergence Selfheating makes convergence harder to reach. The most common problem with convergence is thermal runaway. SPICE sets the junction voltages to large values on the first iteration to ensure the devices are on. With the large forward biases the model equations compute large currents. After the first matrix solution, the returned node voltages tend to be quite large. These factors combine to overpredict the power dissipated in the device, causing an overprediction of the temperature. The overpredicted temperature in the second iteration can cause the temperaturedependent parameters to become large and cause the computed currents to increase enough to cause the temperature to increase even more. The temperature and current will grow geometrically until a numerical overflow occurs, usually after only a few iterations.
PAGE 47
41 Sometimes a variation of this thermal runaway occurs where the simulated temperature becomes high in an iteration and the model overcompensates causing a negative temperature change, then positive again, oscillating with increasing magnitude with similar results as before. Two controls have been put in place to minimize these problems: a hard limit that ensures that only positive temperature changes and a method in which the initial temperatureiteration is skipped. Theses enhancements were made to both implementations to improve the likelihood of convergence by preventing the thermal runaway condition. For the iterationskipping method, the thermal resistance is always held to zero on the first iteration. This is done by checking the initialization flag and skipping the power computations when the flag is set. The zero power from the first iteration results in zero temperature change for the second iteration. The BJT currents and voltages usually settle into the solution region by this time, and the inclusion of selfheating is a much smaller perturbation on the solution. The resulting temperature computed from the power at this point will be closer to the actual value for the BJT, and subsequent iterations usually converge. The BJT model effectively executes two isothermal iterations before the selfheating is turned on. To stop the oscillation a hard temperature limit was added at zero degrees local temperature change that prevents negative temperature
PAGE 48
42 changes. If the model computes a negative temperature change at the beginning of any iteration the temperature change is set to zero for the remainder of that iteration. This allows the model to recover from the oscillation before it becomes a problem, and it prevents endothermic solutions not supported by the GummelPoon model. A third method could be used to constrain runaway conditions in nonlinear equation solutions: stepsize limiting. This method limits the amount a solution element (for example: V35;) can change in a single iteration. This method can work for temperature in the selfheating model, but the best value to use for the maximum step size is difficult to determine and varies greatly from device to device and circuit to circuit. A step size that is too small can slow down convergence noticeably and one that is too large may only slow down runaway but not prevent it. For the BJT model this approach for convergence control gives little return for the computational investment compared to the other two implemented methods, so it is not used in this implementation. It could be useful, however, for special cases, especially if the user had control of the temperature stepsize for each transistor. Results of convergence comparisons between the original isothermal model and the explicit thermal node approach are shown in the next chapter. Such comparisons provide only general estimates of the costs of using selfheating, since convergence in SPICE has seemingly random behavior for large circuits. For example, modifications to the input such as
PAGE 49
43 changing the order in which the elements are listed can affect convergence.
PAGE 50
CHAPTER 4 SELFHEATING IN CIRCUITS SelfHeating Effects An easy way to understand selfheating effects in circuits is through the effects on BJT smallsignal parameters [Miil64, Mill70]. Assume that with selfheating neglected the commonemitter gparameters are given by 8llB=l/^ E (4.1) g 12E = 0 (4.2) S21E = 8mE ( 4 3 ) &22E = 1/r oE (4.4) Using the partial derivatives in Chapter 3, when selfheating is considered, the smallsignal collector current becomes Â• _ g21E v be + g22E v ce + gcgraMjg v be + Ic v ce> (4 g) 1 D C R TH P assuming that i c /i b s I C /I B . Similarly, the base current is j . gllE v be + gl2E v ce + ^B^TH^B^ B v be + ^cc) /AC , b I^bW (4 6) Thus the gparameters, corrected for selfheating, can be expressed as &HE + PbRth 1 ! ( a j\ 8i2B^Pb^thVc g11 1D B R TH P gl2 1D B R TH P (4.8) 44
PAGE 51
45 Â«21 = (4.9) S22 = Â§22E + Pc^H 1 ^ 1 D C R TH P (4.10) The denominators in (4.7) (4.10) become significant as the power approaches the critical values P c = 1/D C R TH or P B = 1/D B R TH where plots of Ic and I3 versus the port voltages have infinite slopes. These limits are related to the onset of second breakdown. In most analog circuits the power dissipation is kept much less than or P B , so that the effects of the denominator are negligible except in circuits requiring great precision. Note that as values for Rp H rise, the power required to cause errors decreases. Now consider g 21 . The electricalonly transconductance g 2 \E * s approximately g mE = I C /V T . If the denominator in (4.9) is close to unity, then the value of g 2 i, corrected for selfheating, is approximately 821^! + d b r th i B V t)Similarly, the electrical only input conductance is She = 1/r nE = Ib /v tÂ» so that 811 = 1/r *= She* 1 + D b r th i b v t)For an y reasonable conditions, the term D B R TH I B V T is very small, so these two parameters are only affected by the corrections in the denominators, which require substantial power dissipation approaching P B . The situation is different for the other two gparameters. A simple model for the electrical output conductance is g 22E = l/r oE = I C /V A , where V A is the Early voltage. From (4.10), for P Â« P c , S22 = Â«22e( 1 +d c r th i c v a)> where the second term in the parentheses can be thought of as a figure of merit giving the fractional error in the output
PAGE 52
46 conductance caused by selfheating for low power. This term can be substantial even for very modest currents. Neglecting selfheating, the effects of changes in V CE on I B , and thus Sl2EÂ» are verv small in forwardactive operation; typically g 12 g can be neglected in circuit analysis. However, including selfheating and assuming the denominator in (4.8) to be near unity, this parameter is much larger in magnitude and positive: 812 = DbRthVc = Â«22E D B R TH I C V A / P<> (4.11) This effect is exploited in the next chapter to extract the thermal impedance of the BJT. Simulation Examples The repercussions of the above results for circuit designers are evident in a variety of circuit types. The following examines the effects of selfheating on some common circuits. The SPICE circuit descriptions for all the circuits shown are provided in Appendix D. All the simulations used the explicit thermal node simulator with thermal impedance subcircuits. The most common analog circuit is the commonemitter amplifier configuration where the emitter and base impedances are typically small and output (collector) load is large, comparable to r 0 . Here selfheating can significantly raise the input resistance and decrease the output resistance. The result is a greatly reduced gain at low to moderate frequencies where selfheating predominates. This effect is illustrated in Figure 41
PAGE 53
47 [Fox93b]. This commonemitter amplifier circuit was originally published by Lee using the implicit temperature simulator developed in this work [Lee92]. The rest of the circuits are new examples of selfheating effects on circuits. The larger 741 opamp circuit shown in Figure 42 is a common twostage amplifier circuit with a low current, low gain differential input stage and a high current, high gain stage [Gra84]. The dc transfer 7K I > Â— i i i inn 1 Â— i i i miii 1 Â— i i 1 1 1 iii 1 Â— i i i ii iii i iii i nn 1 iii m il i 1 1 Â— I'll i i i i m i 1 iii i mi i 1 ' 10 2 10 3 10 4 10 5 10 6 10 7 10 8 Frequency (Hz) Figure 41 Simulated voltage gain vs. frequency for commonemitter amplifier with and without selfheating. Note that the dc gain is halved by selfheating.
PAGE 54
48 characteristics of a 741 opamp circuit are shown in Figure 43 both with and without selfheating. The selfheating curve shows a significant nonlinear distortion. This distortion causes a 10% shift in offset voltage and a large reduction in gain. The greater distortion at low output voltage Figure 42 Schematic for the 741 operational amplifier circuit. Resistance values are as given by Grey and Meyer [Gra84]
PAGE 55
49 indicates the PNP currentmirror in the second amplifying stage, which has no emitter degeneration, has a much reduced output resistance due to selfheating. For the 741 opamp, the first stage is operated at such a low current that selfheating effects are minimal. The second stage is operated at higher currents, and is driven from the low output impedance of an emitter follower, so there is an error in the dc gain due to selfheating. This is shown in Figure 44. When Miller compensation is applied around the 15 1 1 ' 1 1 1 Â» 1 l 0 1 2 Input Voltage (mV) Figure 43 Simulation of dc transfer characteristics of a 741 opamp with and without selfheating.
PAGE 56
50 second stage, the dominant pole frequency is also shifted by selfheating. However, the negative feedback through the compensation capacitor swamps out the positive thermal feedback above the dominant pole frequency, so there is little thermal effect on the phase margin. Some largesignal circuits where precise Vgg matching is important can be affected by selfheating when the collectoremitter voltages of the matched BJTs differ. The BJT with the higher Vqe has greater power 100 Frequency (Hz) Figure 44 Simulation of a 741 opamp openloop voltage gain with and without selfheating showing the effects of compensation.
PAGE 57
51 dissipation, causing its Vg E to decrease to suppress the increase in collector current. These dc effects were demonstrated for current mirrors by Munro and Ye [Mun91], and for current reference circuits by Fox and Lee [Fox93b]. When these circuits are used in switching applications the V BE shifts settle over long time periods as the BJTs heat and cool. Consider the pindriver circuit shown in Figure 45. The function of this circuit is to transfer a voltage level at lowcurrent, high impedance circuit to a highcurrent, lowimpedance circuit (or package pin). Going from V IN to Vqut the signal increases two diode drops through Q 1 and Q 2 Q 3 V OUT Figure 45 Simplified pindriver circuit.
PAGE 58
52 then decreases two diode drops through Q 4 and Q 2 . If all of the transistors are matched the output follows the input. Transistors Q 3 and Q 2 are diodeconnected and do not dissipate much power. QÂ± and Q 4 can have large collectorbase voltages and can dissipate significant power. Figure 46 shows the temperature transients of Qi and Q 4 . For this simulation, the supply voltages Vcc and Vee were 10 V and 10 V, the current sources I x and I 2 were 1 mA and 10 mA, and 50 I Â— Â— 1 Â— 1 Â— ' Â— i Â— ' Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â—  Â— I Â— i Â— i Â— I Â—  Â— i Â— i Â— I Â— r Time ((is) Figure 46 Simulated temperature transients of Q : and Q 4 of the pindriver circuit for and 18 V input pulse.
PAGE 59
53 the input voltage pulse switched from 2 V to 18 V. The temperature swing of Q 4 is much larger than that of The output voltage transients are shown in Figure 47 with and without selfheating. The long thermal transient can be seen in the long output voltage settling time. While the 0.1% settling time for the isothermal case is less than 10 ns, with selfheating the settling time is about 100 \is. The long thermal tail is a problem for highspeed precision applications. 8.05 8.00 7.95 7.80 7.75 7.70 T 1 Without Selfheating With Selfheating Input Pulse i J 1 1 i i L _ i L 0 50 100 150 Time (is) 200 250 Figure 47 Simulated transient output voltage for the pindriver circuit with and without selfheating.
PAGE 60
54 Another application of selfheating in circuit design is thermal coupling of transistors that are near each other. The distance for a device to be considered near to another in thermal terms depends on the circuit. Fukahori and Grey showed that for a operational amplifier the input transistors could be affected by the heat generated in highcurrent output transistors if they are located close to the input transistors [Fuk76]. They showed that careful layout that located the output stage away from the input stage could cure this problem. The general rule in these cases is to lay out highpower elements of the circuit away from the sensitive subcircuits of a design. For the case of multipleemitter BJTs the above rule does not apply. These devices are used in highfrequency applications primarily because of the reduced parasitics from the shorter intrinsic base region and reduced current crowding of the thinner emitter fingers. To simulate a multiemitter BJT with selfheating, it can be separated into multiple devices, one for each finger. This method was developed in collaboration with Brodsky [Bro97b]. A schematic for a threefinger BJT is shown in Figure 48. The fingers are thermally coupled by currentcontrolled voltage sources (Hsources in SPICE). Each finger's thermal subcircuit includes an Hsource for each of the other fingers. An Hsource is controlled by the power flowing in another thermal subcircuit multiplied by the thermal resistance between the two fingers.
PAGE 61
55 Figure 49 shows the result of a dc simulation of the threefinger BJT. The middle finger is heated more than its neighbors because of its close proximity to both of the side fingers. It carries more of the base current and contributes more to the total collector current. The result is a reduced output impedance over the nonthermal simulation of about 70% for this example. Figure 48 Schematic for simulating a threefingeremitter BJT, and the three coupled thermal subcircuits.
PAGE 62
56 The simulation of circuits with selfheating adds to the accuracy of many designs, but is not necessary for all circuits. To further increase the reliability of simulation results, more accurate model parameters are needed, including thermal impedance parameters. A technique for extracting the thermal impedance for BJTs is introduced in the next chapter. The effects of selfheating on the GummelPoon parameter extraction are also examined. 15.0 I 1 1 1 1 1 1 1 1 1 0.0 1 1 1 1 1 Â— 1 i i Â• 0.0 2.0 4.0 6.0 8.0 10.0 CollectorEmitter Voltage (V) Figure 49 Contribution of the middle and side fingers to the total collector current of a threefinger emitter BJT
PAGE 63
67 Convergence Comparison The simulation statistics for each of the above circuits were collected for the selfheating and isothermal cases. The results of a three run average are shown in Table 41. The statistics for the multifinger Table 41 Circuit simulation statistics. Circuit Operatingpoint Normal/Selfheating number/time (s) Ac Normal/Selfheating number/time (s) Transient Normal/Selfheating number/time (s) Commonemitter NPN 5/0.03 6/0.03 46/0.07 46/0.12 741 Opamp 15/0.45 15/0.95 46/1.08 46/1.92 Pindriver 8/0.07 9/0.12 710/2.70 856/5.85 transistor are omitted, because there is not an isothermal equivalent circuit for comparison. These statistics are reported by the ACCT option in SPICE the reports on execution times and iterations for each analysis. The circuits were simulated on a Sun Microsystems, SPARCstation 10 with 64 MB of memory. The computer, at the time of simulation, was only running the FVWM window manager and common UNIX daemons, so the simulation time for each circuit was relatively unaffected by resource demands by other processes. The quality of the computer isolation was shown to be effective, since each of the averaged simulations returned the same result. It is clear that simulations using selfheating have a greater cost in computation time. For the case of the singleBJT commonemitter circuit difference in the operatingpoint solution times is lost in the accuracy of
PAGE 64
58 the measurement, but the selfheating circuit needed an extra iteration implying the need for greater computation time. The ac analysis used 71 percent more time because of the added elements of the thermal impedance subcircuit. Because of the linear nature of the ac analysis, the number of iterations is always exactly the same as the number frequency points computed. For the 741 opamp, the iteration count for the operatingpoint solution is surprisingly the same, but there is a 111 percent increase in computation time for selfheating. It is important not to infer too much from these results, because the bias used for each case was different due to the shift in operatingpoint caused by selfheating. A small bias change for this class of circuits will almost always have an effect on convergence independent of selfheating. The additional time is due to not only the added circuit elements but the increase computation need in the BJT model for selfheating. The ac solution took 78 percent longer, which is consistent with the commonemitter circuit results. The pindriver circuit's initial operatingpoint solution required one more iteration and 71 percent greater time per iteration for selfheating. The effects of selfheating on transient iterations and computation time are expressed in the timestep control. The SPICE algorithm to compute the timestep in the transient analysis increases the size of the time step each iteration until convergence is not reached and then "backs up" to try a smaller timestep. For the isothermal simulation, there are practically
PAGE 65
59 no changes in the output for the majority of the analysis time. SPICE can take large timesteps and complete the simulation with fewer iterations. For the selfheating case the long thermal tail requires smaller timesteps and a greater number of iterations. The time per iteration increased 80 percent. To keep the added overhead of selfheating to a minimum, it is important to use the insight provided in the first part of this chapter to choose the devices most affected for simulating with selfheating. This mixed approach also can improve convergence behavior over using selfheating for every BJT in the circuit. This chapter has used models for thermal impedance for simulation. In the next chapter, the technique used for extraction of the thermal impedance is presented. It was shown here that selfheating changes device performance. The next chapter shows how selfheating can affect the extraction of the GummelPoon model parameters.
PAGE 66
CHAPTER 5 MEASUREMENT OF MODEL PARAMETERS This chapter describes methods for extracting model parameters for use in selfheating circuit simulation. First, a twopart, dc and ac method for extracting the thermal spreading impedance is shown. The results can be used to either verify an impedance model or to directly extract the impedance network for each device [Zwe95a, Zwe96]. Second, the effects of selfheating on GummelPoon dc model parameter extraction are examined. Selfheating causes errors in certain highcurrent parameters if they are extracted under nominally isothermal conditions unless selfheating is accounted for in the optimization phase [Dav88]. Measurement of the Thermal Impedance Two components must be characterized for an accurate model: the thermal spreading impedance (temperature response to a change in dissipated power), and the temperature dependence of the currentvoltage relationships of the transistor. At first glance measuring the temperature dependence of the currents appears straightforward; the ambient temperature is varied by a known amount and the resulting changes in the terminal currents are measured. Actually the situation is more complicated because the actual device temperature rises above the ambient by an unknown amount as a result of selfheating. This unknown 60
PAGE 67
61 temperature change can be significant compared to the ambient temperature change, especially in highcurrent regions. This chapter describes a way to overcome this problem, and a way to use the results to extract the thermal impedance from transient response. Measurement of the base current response is treated as the thermometer in the techniques described here. An underlying assumption of these measurements is that collectorbase conductance from the modulation of the base recombination current is negligible compared to the effect of selfheating on the total base current. If operation is in the forwardactive region where significant impactionization can be avoided, any change in base current resulting from a change in collectorbase voltage with constant baseemitter voltage is dominated by selfheating. Characterization of the thermal impedance requires two steps. The first step involves finding the fractional temperature coefficients: the fractional change in base or collector current for a unit change in the transistor temperature with fixed baseemitter voltage. We find this information by sweeping the dc collector voltage of the transistor at different ambient temperatures and then cancelling the selfheating. The second step is to find the thermal impedance from the base current response to a collector voltage step. Using the temperature coefficient results from the first part, the basecurrent response is converted to a temperature response. After normalizing the transient temperature response by the dissipated power, the thermal impedance model can be fitted to the waveform.
PAGE 68
62 Temperature Coefficient Measurement In this part of the measurement cycle the goal is to calibrate the basecurrent thermometer by measuring the change in current caused by a change in temperature with fixed base voltage. When the temperature is increased the collector current and the power increase, causing the local temperature to increase considerably because of selfheating. An error arises whenever selfheating contributes significantly to the device temperature (at moderate to high current levels). Transient measurements, to be discussed in the next section, are typically made in these regions, and there needs to be a way to characterize the base current dependence on temperature in the presence of selfheating. The total temperature change of a BJT, including selfheating, is given as AT = R TH AP + AT A = R TH (I C AV C + AI C V C ) + AT A (5.1) where AT is the total temperature change at the emitter, AT A is the ambient temperature change, and Rrpn is the thermal resistance. Separately measuring the change in base current caused by a change in ambient temperature and then by a change in power gives two equations that allow cancellation of the selfheating from the temperature coefficients. The power is controlled by varying the collector voltage in the forwardactive region where basecollector conductance is negligible and avoiding regions where impact ionization is significant. When changing the ambient temperature with the collector voltage, V c , constant, (5.1) gives a net temperature change of
PAGE 69
63 at Iav c = o sAT a + r th v c ai c (5.2) where AI C = ^AT = D C I C AT (5.3) is the change in collector current due to a change in temperature. Substituting into (5.2) and solving for AT gives AT A AT Iav c = o = TT5^vc (5 4) When is held constant and Vq is varied, (5.1) reduces to at Iat, = o = RjhC^c + AI c v c) (5.5) where AI c = af AT + aV~ AV c = D c I c AT+ v^c < 5 6 > C A and V A is the Early voltage. Substituting into (5.5) and solving for AT gives I C R TH AV C AT k = 0= 1D C R TH I A C V C Â• (5 ' 7) Notice that the denominators of (5.4) and (5.7) are the same. Combining (3), (5) and the basecurrent dependence given by ai B ai B AI B = ar AT + av^ AV C :sD B I B( AT A + R TH V C AI c) (5.8) where the second term is the basecollector conductance (assumed to be negligible), results in the following ratio:
PAGE 70
64 (AI B /AV C ) AT A = 0 (AI fi /AT A ) = I C R TH (1 + V C /V A ). (5.9) AV c = 0 This can be solved for R^h if V A is known or is much greater than V^. It is important to note that the thermal resistance found here is an effective value, and includes the thermal impedance of the package and die. While this is not an exact value for the intrinsic thermal resistance, it is a good estimate for extremely low frequencies. Once Rth is known, (5.8) can be solved for the fractional temperature coefficient of base current: Note that this rather complicated procedure is not needed at moderate to low currents, where selfheating does not need to be cancelled. The following is a description of the measurement setup and algorithm. The transistor is connected in commonemitter configuration inside an oven or on a probe station thermal chuck. It is important to have stable, accurate ambient temperature control during the measurement, because the voltage sweep takes a long time and errors in the ambient temperature significantly affect the accuracy of the results. The algorithm for the measurement is simple, though very timeconsuming. First, set the ambient temperature and allow it to settle completely, chose a base voltage for the region of interest (moderate to AT A + R TH V C AI C (5.10) AV c = 0
PAGE 71
65 high current region), and sweep the collector voltage in the forwardactive region where a noticeable, linear increase in I B can be seen (avoiding any falloff due to impactionization). Use a very long delay time between measured points; about 3060 s typically works well. This allows the selfheating of the transistor to settle before making the measurement at each point. Execute the sweep, measuring 1q, Ib> an d Vce fÂ° r each Vbe of interest. Change the temperature and repeat the voltage sweeps done before (from experience, a temperature step of less than 10 Â°C is recommended to avoid errors caused by the strongly nonlinear dependence of current on temperature). Take the average current changes with respect to collector voltage and ambient temperature. Using a geometric mean helps compensate for the nonlinearity of the data. Use (5.9) and (5.10) above to compute the fractional temperature coefficients at each base voltage value. Figure 51 shows a plot of the temperature coefficients versus base voltage for an NPN Harris UHF transistor with a 3x50 im 2 emitter. The UHF process yields highfrequency, dielectrically isolated, vertical transistors [Dav92]. The collector voltage was swept from 24 V while the base voltage was varied from 700850 mV in 5 mV intervals using a HewletPackard 4145A Semiconductor Parameter Analyzer. External computer control was used to run the HP 4145A, because only short durations are available for the voltage steps using the integrated controller. The automation of the experiment also helped reduce human
PAGE 72
66 errors and total measurement time by integrating setup, measurement, and data collection into a single routine. The ambient temperature was controlled using a Delta Design Model 111 Thermal Chamber; temperatures used were 25 and 33 Â°C. This measurement has errors that are dominated by oventemperature accuracy, nonlinearity of the data, and effects of base and emitter resistances. The thermal chamber has a published repeatability t 1 1 1 1 1 1 1 1 1 1 1 1 r 0.70 0.75 0.80 0.85 Base Voltage (V) Figure 51 Harris 3x50^im UHFN3 bipolar transistor fractional temperature coefficients of base and collector currents versus baseemitter voltage measured with and without correcting for selfheating.
PAGE 73
67 accuracy of 0.5 Â°C and a relative error of 0.1 Â°C. The absolute error is not as important for the temperature coefficient measurement as the relative error since the change in temperature is used in the computations. Because the measurements shown here were repeatable with an error consistent with the 0.1 Â°C, the published errors can be believed. The emitter and base resistances cause a slight error in the intrinsic V BE for high currents. When AI C is 1 mA the intrinsic AV BE is reduced by 2 mV. The total error for the experiment is approximately 1 0 percent. Thermal Impedance Measurement The objective of this part of the measurement cycle is to measure the normalized thermal response to a step in power. This is done by converting the base current response into a temperature response using the temperature coefficients extracted above. Solving the fractional temperature coefficient definition for the change in temperature gives where Iq is the median value of the base current for the transient. At this point the signal represents the thermal transient for the power step applied, and it must be normalized for a unit power step. This is done by dividing the AT(t) signal by the power of the step: AT(t) = I B (t) I B (0) r B D B (5.11) AT (0 = AT(t) (5.12) norm AI C V C + AV C I C where the denominator term is the magnitude of the power step.
PAGE 74
68 The setup for this measurement is similar to the temperature coefficient commonemitter configuration with a collector voltage step applied instead of a slow voltage sweep. The setup is shown in Figure 52, where the BJT in the center is the test device being probed on a wafer. The currenttovoltage converter transforms the base current into a voltage waveform for the oscilloscope. With the base voltage held constant, the base current is V<) = (Vqut v be) /r f (5.13) where is the feedback resistance. Unfortunately, the intrinsic baseemitter voltage is not constant because of the parasitic base and emitter resistances. Since the change is small compared to the thermal voltage, a linear correction is suitable: WO = V QUT ~ V BE R F R B (1 + P)R E (5.14) where Rg and R^ are estimated base and emitter resistances and (3 is the largesignal current gain. R S =50Q V c Pulse Generator ^3R F =51(K>
PAGE 75
69 There are important practical tradeoffs in choosing the feedback resistor value. The most important is the gain versus bandwidth tradeoff. Gain must be large because the signals of interest are typically very small, yet a wide bandwidth of about 100 MHz must maintained to resolve the earliest part of the transient response. In our measurements, an amplifier with gainbandwidth of 1 GHz was used with R F = 510 Q and RscOPE = 50Q for a loop gain of about 1 0 and a bandwidth of 1 00 MHz. The "sponges" shown in Figure 52 are highfrequency sinks that absorb the energy at frequencies beyond those of interest in the thermal measurement; they are required to keep the UHF transistors from oscillating. The sponges also help absorb some of reflected power from the unmatched impedances of the pulse generator and BJT collector. The device under test was probed onwafer using one meter, 50 Q. coaxial cables. The following is a description of the method used to make the stepresponse measurement. First, a value of base voltage was selected to give a current large enough that the collector voltage step created enough power to get a measurable change in base current for a square wave collector voltage of 25 V. The best frequency to use for the square wave depends on the transistor; the response must settle before the end of each half cycle. Much of the response occurs in the first 50 us, but it continues until about 1 ms, after which package heating dominates. Because of the distributed nature of the response, the opamp output was recorded using three overlapping time windows of increasing size to capture with high
PAGE 76
70 resolution the fast beginning of the response while keeping the number of data points to a minimum for the long tail. Next, the pulse generator was set to dc output and an ammeter was connected in series with the collector to measure the collector current at the pulse maximum and minimum values. The opamp output is also measured at this time to compute the base current from (5.14). From this data the (3 of the transistor is determined, so that the transient base current can be converted into a collector current for computation of the power step magnitude. The base current minimum and maximum in this dc measurement differ from the values in the transient because of the difference in mean power being dissipated. Using (5.11) and (5.12), the normalized temperature response can be computed. Figure 53 shows the normalized temperature responses for a range of geometries. Figure 54 shows the thermal resistance of the same transistors versus emitter width. Estimated errors for the measurement are also shown. These errors are primarily from three sources: errors in the extracted fractional temperature coefficient, the variance of (3 over the range of collector voltage, and the absolute error of the measured base current. The absolute error of the base current seriously limits accuracy for small transistors that operate at low currents where it is difficult to get a large enough change in I B to be seen over the noise of the measurement equipment. This does not mean that thermal effects in these transistors
PAGE 77
71 are insignificant. In the setup used here the limit for dc collector current is approximately 1 mA. For a small transistor biased at 1 mA undergoing a 5 mW power step with an R TH of 500 K/W, an approximately 15% increase in I B occurs. However, this change is seen as only 1 mV at the opamp output. This signal is hard to resolve with a oscilloscope with a 1 mV/ division minimum resolution. Add to that a typical 2 mV of noise superimposed on this signal, and it becomes very difficult to resolve the transient even with averaging. 400 350 300 Â« 250 8 200 B U H 8 150 N Â— E 100 c o Z 50 01 Â— i i i GÂ— 03x30 Â— B3x50 3 Â— 03x70 AÂ— A3x90 < Â— <3xll0 VÂ— V3xl70 > Â— Â£>3x210 i i 1 1 1 1 1 i i Â— i i 1 1 1 1 50 1010 6 10" 5 Time [s] 10" 10 3 Figure 53 Normalized temperature response versus time of the UHF transistor for seven different emitter widths.
PAGE 78
72 Thermal Impedance Model and SPICE The above results must be compressed into a manageable model for use in circuit simulation. Using the version of SPICE modified to include a temperature (AT) terminal described in the previous chapter, the thermal impedance can be incorporated into precision circuit simulation of dc, ac and transient. To convert the measured stepresponse data to an equivalent RCladder model, the program TIPP was used to provide SPICE with calculated values for the thermal resistance and thermal capacitance components needed for the thermal equivalent circuit [Bro93]. TIPP can handle up to five poles for the thermalimpedance model, and input can come from either measured transient data (AT vs. time as found 400 0 1 Â— 1 Â— 1 Â— 1 Â— 1 Â— 1 Â— Â• Â— 1 Â— ~ Â— Â— ' Â— Â— Â— Â— Â— i Â— Â— Â— . Â— . Â— i . , . J 0 50 100 150 200 250 Emitter Width [!m] Figure 54 Thermal resistance versus emitter width for the UHF transistor with emitter length of 3 (im
PAGE 79
73 here) or a predictive physical model [Joy70]. Figure 55 shows the normalized temperature response for a 3x70 im 2 transistor and the resulting threepole and singlepole fits. The threepole fit is good, and generally is the best tradeoff between accuracy and complexity for precision applications. For some situations the singlepole model is sufficient. Many issues must be considered when deciding whether to use measurement or a physicsbased predictive model for determining the thermal impedance in a process. Typically, manufacturers extract model parameters for each transistor in a process from measurement. The Time [s] Figure 55 Comparison of normalized temperature response for one and two pole fits and measured data of the UHF 3x70 im transistor
PAGE 80
74 extraction is automated and can be done fairly quickly. If the measurement of the thermal impedance is included in the extraction, the time spent on each transistor is greatly increased. Consider just the calibration step of extracting the current temperature coefficients. Estimating that each data point takes 30 s and that the minimal of two Vqe values per Vg^ (which could case large errors) for two temperatures, gives a minimum of 2 minutes per V BE . Add to that the time for the change of temperature of approximately 10 min and the time to find the correct operating region for the measurement. This takes about 30 min per transistor before the transient measurement has been made. Considering the time that would be required to measure parameters for every BJT geometry, development of a model for thermal impedance dependent on process parameters that scales with transistor geometry would be preferable. Such a model should use physics as much as possible but exclusively. The tradeoff here is the accuracy. A physical model cannot include all effects and still yield results quickly. For each neglected effect the accuracy usually decreases. For some structures the assumptions needed to solve the heatflow equations may not allow enough accuracy to be useful. For these structures the loss of accuracy in the physics can usually be corrected by using empiricism and the information from measuring only a few transistors. In the case of a vertical, junctionisolated, rectangular emitter BJT good predictive physicsbased models
PAGE 81
75 exist (such as the one described in Chapter 2). For the waferbonded UHF transistors measured here, a scalable thermal model is difficult to develop due to the large process variation of the oxide depth that greatly affects the thermal impedance. For thin SOI technologies where the isolation depth and thickness are wellcontrolled, physical modeling seems simple until the increased importance of heatflow out the contacts and wires is considered. Extraction of thermal impedance for some transistors in a process will always be needed either to validate a model or to tune some empirical parameters. A reasonable goal is to avoid spending needlessly long times measuring every BJT in a process. Selfheating and Model Parameter Extraction When standard parameterextraction procedures are used to find parameters such as Early voltage and knee current, the data typically include some selfheating effects. When these parameters are used in a model with selfheating, the selfheating effects are overrepresented. For accurate modeling, the parameter extraction procedures themselves must be modified to separate thermal and electrical effects. This section describes a process to accurately extract these highcurrent parameters. Most of the SPICE GummelPoon model parameters are normally extracted in lowpower regions where selfheating effects can be ignored. Three parameters, however, are usually extracted in highpower regions where selfheating can be important. These are the knee current, the Early voltage, and the parasitic base resistance (R B ). I KF and R B are
PAGE 82
76 extracted from a Gummel plot at high current, usually with a significant basecollector voltage to avoid quasisaturation, so that power dissipation may not be negligible. V A is extracted from the slope of a collector current versus collectoremitter voltage plot, which is very sensitive to the power dissipation of the measured device. It is tempting to try pulse measurements to avoid selfheatinginduced problems by measuring the device before it can heat up. Unfortunately, because the thermal response in scaled devices has time constants on the order of nanoseconds, it is very difficult to measure a BJT in electrical steadystate before significant selfheating begins. In the approach presented here, the device is allowed to reach thermal steady state before data is recorded. For comparison, a normal isothermal extraction not accounting for selfheating was done first. The model parameters were extracted and optimized in the normal way for a packaged 1.6 x 5 (im) 2 , junction isolated, NPN transistor obtained from Analog Devices [Get78]. All measurements were made using a temperaturecontrolled oven to supply a constant ambient temperature. The data for the characteristic curves shown in Figure 56 were measured from a 03 V sweep of the collectoremitter voltage with the base voltage held constant at 0.8, 0.85, and 0.9 V. Each point is the average of 16 measurements made after holding the bias for 20 seconds so the temperature could stabilize. The same delayandmeasure technique was used to record the Gummel plots used in the extraction. For the Gummel plots the baseemitter voltage was swept from
PAGE 83
77 0.51.1 V with the collectorbase voltage held constant at 0.5 V to avoid quasisaturation. Note that it would not be uncommon to use a V^b of 1 V, doubling the power dissipation seen here. After the device Gummel plots and characteristic plots were measured at several temperatures, the initial parameter set was extracted using the curvefitting techniques described by Getreu [Get78]. The exception to this was the base and emitter resistances were extracted using the method described by Ning and Tang [Nin84]. This is an adequate 4.0 i Â» Â» ' i i i i < ' i ' i Â» i 0 0 1 1 ' ' 1 1 1 I I 1 i_ 1 ' 0.0 1.0 2.0 3.0 CollectorEmitter Voltage (V) Figure 56 Characteristic curves resulting from a normal model parameter extraction.
PAGE 84
78 method for getting values for these parasitic parameters when ac methods are unavailable. The extraction was for only the parameters needed for simulation of the forwardactive and saturation regions; the reverse modes of operation were ignored because of their infrequent use in circuit design and the obvious extension from the forward modes examined here. Next the model parameters were tuned using nonlinear optimization, the extracted data, and the output of SPICE simulations of the measurement using the parameters. Most of the parameters required only small corrections with the exception of the Early Voltage. Results of the optimization are shown in Figure 56. This simulation result is a poor fit to the measured data especially at high currents where the effects of selfheating can be seen in the increased slope of the Iq versus V^e curve. One way to correct for such selfheating errors is to measure the thermal resistance and use it with the self heating model in SPICE. The extracted effective thermal resistance using (5.9) was 800 K/W. To complete the selfheating model, the SPICE thermal parameters XTI and XTB were also extracted using the measured data over many temperatures. The resulting fits to the data are shown in Figure 57 and Figure 58. When selfheating is added to the simulation using the same model parameters, as shown in Figure 59, the conductance modeling is improved but the accuracy of the current modeling is not. It is not enough to simply include selfheating after an isothermal extraction.
PAGE 85
79 The model parameters were optimized once more, this time including selfheating in the model. The Early voltage was not optimized; the value computed from (5.9) during the thermal resistance extraction was used instead. The results are shown in Figure 510. The fit to the data is much improved with good agreement in both slope and magnitude of the collector current. The model parameter optimization including selfheating only affected three parameters significantly: V A , and RbThe value of Ij^p 10 12 i Â— i Â— i Â— i  i Â— i Â— i Â— i  i i i Â— iÂ—  Â— r Â™i 1 1 V 20 _i i i i_ O Measured Data Â— Model J L. 30 40 50 60 Temperature (Â°C) 70 80 Figure 57 Resulting model fit for extraction of XTI from the Ig versus temperature data.
PAGE 86
80 in this optimization was 40% lower than that found in the isothermal extraction. Selfheating increases the collector current in the isothermal extraction, so the electrical value of kneecurrent is overestimated. A 50% increase in the electrical value V A results from taking into account the increased slope of the characteristic curves due to selfheating. Finally, a 25% decrease in the base resistance is likely caused by temperature dependence of the current gain. Although the change in base resistance 1 80 i Â— ' Â— Â— i i i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— i Â— O 140 20 30 40 50 Temperature (Â°C) j i , L 70 80 Figure 58 Resulting model fit for extraction of XTB from the versus temperature data.
PAGE 87
81 when selfheating is included, it is likely that R B is best extracted using highfrequency ac methods [Get78]. CollectorEmitter Voltage (V) Figure 59 Characteristic curves resulting from the same model parameters with selfheating added to the simulation.
PAGE 88
82 Figure 510 Characteristic curves resulting from including selfheating in the extraction process.
PAGE 89
CHAPTER 6 SELFHEATING FOR SOISPICE SOISPICE is a simulation tool developed at the University of Florida to do physicsbased modeling of silicononinsulator metaloxidesilicon fieldeffect transistors, SOI MOSFETs, for circuit simulation [Fos96, Suh95, Yeh95]. The basic structure for these devices is shown in Figure 61. There are two variations of this device, differing in the thickness of the body layer. In one case the body layer is thick enough so that in normal saturation region operation the depletion region does not reach the bottom of the body, and in the other case the body is thin enough for the depletion region to reach the bottom. These structures are called either partially depleted, PD, or fully depleted, FD, SOI MOSFETs. Source Gate Drain Back Gate / Substrate I I Silicon Q Silicon Dioxide Q Polysilicon Figure 61 Basic structure of a SOIMOSFET with Si02 insulating layer. 83
PAGE 90
84 The SOI MOSFET device is surrounded on all sides with an insulating layer, usually Si0 2 , that forms a barrier to heat flow. SOI devices tend to get much hotter than their bulksilicon counterparts, because the heat is trapped in the small silicon island the device is built in [Goo95]. The dominant temperaturedependent effect on MOSFET current is the carrier mobility, which is a much weaker function of temperature than the dominant effect (carrier concentration) on BJT current. For comparable thermal impedances and power dissipations, selfheating can be ignored in a bulk MOSFET when it cannot in a BJT. Because of the insulation layer surrounding a SOI MOSFET, the thermal impedance can be one or two orders of magnitude greater than that of a comparable geometry bulk device. This increase in thermal impedance causes generation of much greater temperatures within the SOI MOSFET, and even the weak temperature dependencies of the transistor become significant. SelfHeating Implementation The following shows how the selfheating model was implemented into SOISPICE, and then some simple circuit simulation examples are shown. The implementation has many similarities to the BJT approach, but some major differences exist. Some of the differences result from the organization of SOISPICE. Other differences were requested by SOISPICE users, who preferred a singlepole thermal impedance, implemented internally, without additional inputfile circuit elements, as
PAGE 91
85 was done in the BJT implementation. The organization of the SOI MOSFET model allows all of the current and voltage derivatives to be easily computed numerically. This numerical approach was a necessity, because few of the physical model equations have closedform solutions for their derivatives. The following discussion will focus on these features that were unique to the SOISPICE selfheating implementation. The details of the SOI MOSFET electrical modeling are still under development and are beyond the scope of this work, so they will be ignored when possible. Detailed FORTRAN code modifications to SOISPICE for this selfheating model can be found in Appendix C. The SOI MOSFET Model For reference, the schematic for the SOI MOSFET model is shown in Figure 62. The model has many current sources, but most are reactive and dissipate no real power. The dominant sources of real power are the channel current, Ic H , and the parasitic BJT current, Igjp The other real current sources, I R Q t and Iqj are orders of magnitude smaller and do not contribute significantly to the power dissipation. Other sources of power dissipation include the parasitic resistances, specifically the drain and source resistances. These may not always be negligible in computing the total power dissipation, but a problem arises with the thermal impedance modeling if they are included. The thermal impedance modeling assumes the power is generated in the channel region under the gate, and not in the large adjacent source and drain regions. However, since this
PAGE 92
86 implementation uses a simple, less accurate, singlepole model, including the source and drain resistances in the power computation adds more to the total accuracy than the thermal impedance modeling deficiency takes away. Therefore, these resistive power sources were included. Specifying SelfHeating The thermal subcircuit for the SOI MOFET is shown in Figure 63. The ground connection, T A , represents the ambient temperature of the circuit defined globally in the input file. No external connections are !ch Figure 62 Schematic for the SOISPICE model [Fos96].
PAGE 93
87 allowed to the thermal circuit in this implementation so the device description line becomes: Zxxxx nd ngf ns ngb [nb] [BJT] model RTH= CTH= L= W= ... where RTH and CTH are two new deviceline parameters that describe the thermal subcircuit. If the parameter RTH is not specified on the device fine, the thermal subcircuit is not generated, and the standard isothermal model is used. Since there are no external connections to the thermal subcircuit and the ground node is already defined, only one new node is needed to define the temperature in the selfheating model, so only one new row and column are added to the device matrix. Computing the Derivatives SOISPICE computes all of the current and charge derivatives with respect to node voltages numerically. The currents are computed from the model equations once at the bias point of the current iteration. For each node voltage the model is recomputed with a perturbation, and a finitedifference approximation to the derivative with respect the node voltage is computed. For selfheating this process is extended to include another Source Figure 63 The thermal subcircuit added to the SOIMOSFET model for simulating selfheating.
PAGE 94
88 perturbation computation for the temperature, and the power is added to the list of currents. For this process to have the desired results, the model equations must include a power computation and the equations for the model parameter temperature dependence. Computing the Power Dissipation The equation used for the power dissipation is p = I^h + ^jtIVcIs + ^ch + ^jtI^Rd + Rs + Rldd + Rlds)' < 6 D where Iqh is the channel current, Ibjt is the parasitic BJT collector current, and Rg are the drain and source resistances, and Rldd Rlds are the low doped drain and source region resistances. The other modeled currents that are not included in this power dissipation computation, the thermal generation and impactionization currents, flow in regions away from the channel and are an order of magnitude smaller than the two included currents. Local Temperature Update of Parameters The second task of including the temperature dependence of the model parameters is achieved using a new subroutine. This subroutine contains all of the SOI MOSFET temperature equations found in the global temperature update routines. These equations have been optimized to reduce computation time. The parameters affected are shown in Table 61 with a short description of each one. These dependencies were implemented for the SOISPICE global temperature model by Glen Workman [Wor96]. Equations for both the partially depleted and fully depleted devices are included in the subroutine. A flag is used to select between
PAGE 95
89 Table 61 List of model parameters with temperature dependence. Pq i*o rn of oÂ»< X ciL CtHlC lei Name XNIN Intrinsic carrier density EGFET Silicon energy gap VTT Local Thermal Voltage XMULDSJ LDD carrier mobility XMUBH Minority carrier mobility in highdoped body UO Zero field electron mobility VSAT Carrier saturated drift velocity WKF Frontgate work function difference WKB Backgate work function difference VFBF Frontgate flatband voltage VFBB Backgate flatband voltage JRO Bodysource junction recombination current coefficient ALPHA Impactionization parameter a 0 TAUG Carrier lifetime in lightly doped region VBI Source to bulk builtin potential PHIB Twice Fermi potential of body RLDD LDD sheet resistivity BETA P FSAT fsat equations that the PD and FD devices do not have in common. The result is that only one routine must be modified if the temperature equations are modified in the future. As was the case in the BJT selfheating implementation, the thermal voltage must be treated with special care. This parameter is defined globally for all devices in SPICE, and a local
PAGE 96
90 value (shown as VTT in Table 61) must be denned for the model equations. Loading the Jacobian The loading of the selfheating elements into the Jacobian matrix is done methodically. The current derivatives with respect to temperature are loaded in the new temperature column at the rows corresponding to the current branch nodes with the sign indicating direction of flow. The power derivatives with respect to voltage are loaded in the temperature row at the columns corresponding to the two nodes that comprise each voltage. The thermal conductance is loaded at the temperature diagonal element. The terms associated with the thermal capacitance are not loaded at this point, but instead are handled during preprocessing. The Jacobian matrix load for the ac analysis is implemented in nearly the same way, using complex variables for the reactive elements instead of the currentconductance pairs of the transient analysis. The simplified flowchart of how the thermal model implementation works is shown in Figure 64. SOISPICE already has the capability to deal with linear capacitors. It is a burden on the programer to implement all of the state variables and computations to provide for the thermal capacitance inside the SOI MOSFET routines. By developing a way to trick SOISPICE into believing that the thermal capacitance is the same as any other capacitance, SOISPICE processes the thermal capacitance with all of the other circuit
PAGE 97
91 linear capacitor elements. This is done by adding a feature to the SOISPICE FIND subroutine. The last argument of the subroutine call, IFORCE, is used either to return the location of a device in memory or create space in memory and store a device there. The latter option checks whether the device has an unique name and returns an error if it does not. A third option was added that stores the device without checking the name uniqueness. Using this option the thermal capacitors are all added to the READIN MODCHK TMPUPD I NFDFET Decode device line including RTH and CTH Check if thermal model is needed. 1 . If so, generate thermal subcircuit. 1 FIND Add CTH to the list of circuit capacitors. Update model parameters for ambient temperature. Start Newton loop 1 NFDID Compute derivatives for the Jacobian including temperature. 1 NFDMOD Evaluate model at local temperature. SOITMP Done New routine to update model parameters for local temperature Figure 64 Flowchart of SIOSPICE execution showing processing done for selfheating.
PAGE 98
92 bottom of the normal capacitor list with the proper connections to their thermal subcircuits, all having the same name to avoid possible name collisions with the regular circuit capacitors. SOISPICE then processes the thermal capacitors like any other linear capacitor in the circuit without adding more code to the SOI MOSFET model routines. This approach also improves efficiency. Since any computations added to the model routines are executed twice per iteration to compute the partial derivatives, excluding the computations associated with thermal capacitance from the model routines reduces computational overhead. Convergence Issues Because of floatingbody effects and other complex SOI MOSFET effects, convergence control and initial conditions must be set with care even without selfheating to achieve convergence of some circuits. Selfheating tends to degrade the robustness of convergence, but this is expected any time the dimension of the Jacobian is increased as it is for selfheating. The difficulty of estimating the extent of the degradation is exacerbated by the convergence behavior of the isothermal model, which was being refined concurrently with the selfheating implementation. Either the technique of skipping the first Newton iteration as used in the BJT implementation or a traditional temperature stepsize limiting algorithm could be incorporated to enhance convergence.
PAGE 99
93 Simulation Results Three example circuits were simulated to illustrate some of the effects of selfheating. The first is a dc simulation to generate the characteristic curves. The commonsource configuration is used with the sourcedrain voltage swept from 0 to 3 V for three different gate voltages. The results for the floating body case are shown in Figure 65. 5.0 DrainSource Voltage (V) Figure 65 DCsweep simulation for an nchannel, floatingbody, PD, SOI MOSFET with 0.2 urn length and 10 um width. V DS was swept from 0 to 3 V for V G of 1, 1.5 and 2 V. The thermal resistance was 5 kW/K
PAGE 100
94 The second is a stepresponse test of a single nchannel, PD, SOI MOSFET. Again, the commonsource configuration is used with the gate biased for the saturation region. The drain voltage is then pulsed. The resulting drain current response is shown in Figure 66. Unlike the BJT behavior, the current decreases from the effects of selfheating. The dominant temperaturedependent carrier mobility decreases with temperature and the drain current along with it. Other than the obvious 4.5 I > 1 > 1 1 1 1 1 1 Selfheating No Selfheating J J Q 3.5 3.0 0.0 0.1 0.2 0.3 0.4 Time ((is) 0.5 Figure 66 Transient current response to a drain voltage step for a commonsource, nchannel, 0.2x1 Oum, PD, SOI MOSFET. Vg=2V Vds=23V Rth=5kW/K
PAGE 101
95 biasing errors that can result, this negative temperature coefficient can cause hysteresis in the dc transfer characteristic when the thermal resistancedrain current product reaches a critical value [Fox93a]. The third example demonstrates that selfheating is not a significant problem for complementary MOS logic circuits; the primary application for SOI MOSFETs currently. The circuit for a clocked complementary SOI MOSFET inverter is shown in Figure 67. Figure 68 shows the output voltage waveform and the temperature waveform of the nchannel device. Notice that the temperature increases in steps synchronized with the switching of the inverter. The only time a complementary gate dissipates significant power is during switching. This is shown as a steep step in the temperature waveform, and a slow cooling until the next transition. The simulation shows the temperature Figure 67 Complementary inverter circuit used for dynamic power dissipation example. Vi N is pulsed from 03 V at a frequency of 40 MHz and V DD is 3 V.
PAGE 102
96 approaching a steadystate value. The final value of the temperature is a function of the total dynamic power dissipation for the inverter: ,2 P D = c L (v DD ) (6.2) where C L is the load capacitance, V DD is the gate bias voltage, and t is the period of the clocking signal. For the inverter used here, ,2 _ 0.5pF(3V) = 25 ns (6.3) 4.0 > 3.0 sr +^ iÂ— i 2.0 1.0 f o 0.0 u & 0.6 u 8 I 1.0 Time ((is) Figure 68 Inverter output (top) and temperature change (bottom).
PAGE 103
97 Since Pq is the total power for the inverter, half of the power is dissipated in each transistor. Multiplying half the result of (6.3) by the thermal resistance of 5 kK/W gives about 0.5 K increase in temperature, which is consistent with the simulation steadystate temperature. Such a temperature increase is too small to significantly affect the circuit performance, which is highly typical of selfheating in CMOS digital circuits. This is not to say that thermal effects cannot be important in digital circuits. The heat generated must ultimately leave the package, and for large, dense circuits the thermal impedance of the dietopackagetoambient is important for reliability, since it controls the rise in average temperature. However, selfheating is usually insignificant, since the instantaneous device temperature does not follow the instantaneous power dissipation. Conclusion A selfheating implementation was shown for the SOI MOSFET models in SOISPICE. A singlepole thermal subcircuit was added to model the power dissipation through the thermal impedance of the device. Three example simulations were shown. The step response showed how the negative temperature coefficient of the current affects biasing and other analog applications. The inverter simulation demonstrated CMOS logic circuitry where the contribution of selfheating to the device temperature is insignificant. The greatest impact of selfheating in SOI MOSFET circuits will likely be on analog circuits and in parameter extraction.
PAGE 104
CHAPTER 7 CONCLUSION An extensive study of selfheating modeling and effects has been presented here. A historical look at selfheating for circuit simulation showed some continuing deficiencies in modeling. The quality of the modeling has been improved by this work in several areas. Three implementations of selfheating in the popular circuit simulator SPICE were shown: two for BJTs and one for PD and FD, SOI MOSFETs. Circuit effects of selfheating were examined for dc, ac, and transient analysis. A new measurement technique for extraction of thermal parameters was presented. The two approaches to implementing selfheating in the GummelPoon, BJT model each had their advantages. The implicit temperature approach did not require any changes to the Jacobian matrix. This approach works well for dc and ac analyses, allowing physically accurate modeling of the thermal impedance. The inability of this implementation to simulate transient analyses necessitated finding a different approach that could. The second approach enabled transient analysis by treating the power dissipation as an electrical analog. The power and thermal impedance formed a thermal subcircuit that allowed explicit computation of a temperature increase. The thermal subcircuit required extending the 98
PAGE 105
99 device matrix to include the explicit temperature node. Because the thermal impedance is modeled with an RC network, a tradeoff between accuracy and complexity was required. The circuit application determines the level of accuracy needed. The last implementation for the SOISPICE models PD and FD, SOI MOSFETs used a singlepole thermal subcircuit model to model selfheating. The simplicity of the thermal model is reflected in its easy use, requiring only two new parameters on the device line. A numerical approach to computing the derivatives for all of the elements of the Jacobian made the addition the matrix node for temperature easier than the previous implementations for BJTs. An examination of the effects of selfheating on circuit performance was presented that included techniques for analyzing these effects in hand calculations. Simulations of a commonemitter amplifier showed the ac and dc effects on openloop gain and the transfer function distortion for BJTs. The pindriver circuit demonstrated how long thermal settling times can occur in switching applications. The SOI MOSFET simulations showed that care must be used in analog circuits and that digital applications are not directly affected by selfheating. A new measurement technique for extracting the thermal impedance of BJTs was shown. The twostep procedure used a dc calibration step to extract the current temperature coefficients, and a highspeed measurement of the base current step response to extract the
PAGE 106
100 temperature waveform. The results of the first step are then used to scale the base current into temperature. After normalizing for power, the temperature response is fitted with a multiplepole model for use in the thermal subcircuit. The effects of selfheating on GummelPoon model parameter extraction was examined by including the selfheating model in the optimization simulation. The Early voltage, knee current, and base resistance were shown to be the parameters affected most by selfheating. Selfheating is a potentially important effect that all circuit designers should be conscious of in their designs. It can have devastating effects on some precision circuits if not accounted for before manufacturing. Selfheating is not an important factor in every circuit, so designers should use discretion when using it in simulation. There are penalties in simulation time and convergence if selfheating is applied carelessly.
PAGE 107
APPENDIX A SPICE 2G.6 CODE CHANGES FOR MODELING OF SELFHEATING USING THE IMPICIT TEMPERATURE APPROACH The following is a summary of the changes made to SPICE 2G.6 to implement selfheating in the BJT model. The local temperature of each device is computed for each iteration as a function of the power dissipation in the collectorbase spacecharge region (SCR). Temperature is not treated as a state variable, and does not require addition of a new row and column in the circuit matrix for each BJT. Data structure expansion for new model parameters The SPICE data structure must be expanded to include the new model parameters, state variables, and device variables necessary for the thermal model. A subset of the following six new model parameters are needed to specify the thermal model shown in Table Al. Table Al Added model paramters. RTH Thermal resistance CTH Thermal capacitance NEPI Epilayer doping concentration LTH Emitter length WTH Emitter width DTH Collectorbase junction depth 101
PAGE 108
102 RTH and CTH combine for the singlepole model for the thermal impedance, ZTH, or ZTH can be computed from RTH using a frequencydomain model for a point heat source. The last four parameters can be used to predict RTH from the device bias, doping, and geometry, or RTH can be supplied directely. To simplifiy future modifications, eight new model parameter positions were created, with the two extra parameters left undefined. Two variables were added to the state vector shown in Table A2. Table A2 Statevaraiables. TEMPO Device local temperature change RTHO Device thermal resistance Since these are computed "on the fly" it is helpful to know what their values were in the previous iteration, especially in dc sweeps. Six variables were added to the device shown in Table A3. Table A3 Device variables. GPIE electricalonly baseemitter conductance GMUE electricalonly basecollector conductance. GME electricalonly basecollector transconductance. GOE electricalonly collectoremitter conductance. TCB temperature coefficient of base current. TCC temperature coefficient of collector current. These operatingpoint values are used in the ac analysis.
PAGE 109
103 Changes to the Preprocessing and Postprocessing Routines First the data structure must be expanded to include the parameters described above: eight modelcard variables, two state variables, and six device variables. Subroutine ADDELT This routine adds an element and defines the size of linkedlist arrays for computation of offsets. Add six device variables and eight model variables. DATA LVAL(12) = 4 LVAL(12) = 10 LVAL(22) = 55* LVAL(22) = 63 Subroutine ALTER.f This routine changes the elements or device parameters for multiple SPICE runs. Add six device variables and eight model variables. DATA LVAL(1 2) = 4 LVAL(1 2) = 1 0 LVAL(22) = 55 Â» LVAL(22) = 63 Subroutine DCOP.f This routine prints out operatingpoint information. The following changes allow thermal resistance and local temperature to be printed in response to the .OP command. DIMENSION 2 CCS(1 2), CBX(1 2), RX(1 2),TEMP(1 2), RTH(1 2) DATA ATEMP, ARTH / 4HTEMR 3HRTH / Change BJT section to: BETADC(KNTR)=CC(KNTR)/DSIGN(DMAX1 (DABS(CB(KNTR)),1 OD20),
PAGE 110
104 1 CB(KNTR)) TEMP(KNTR)= VALUE(LOCT+1 9) RTH(KNTR)=VALUE(LOCT+20) IF (MODEDC.NE.1 ) GO TO 325 IF (VALUE(LOCT+20) EQ. 0.0D0) GO TO 355 WRITE (IOFILE.AFMT2) ATEMP,(TEMP(I),I=1,KNTR) WRITE (IOFILE.AFMT2) ARTH,(RTH(I),I=1 ,KNTR) 355 IF (MODEDC.NE.1) GO TO 360 These thermal variables are printed only if a thermal model is used. Subroutine FIND.f This routine presets storage for each input element. Add six device variables and eight model variables. DATA LVAL(1 2) = 4 + LVAL(1 2) = 1 0 LVAL(22) = 55 + LVAL(22) = 63 Subroutine MATPTR.f This routine builds the sparse matrix. The offset for the LXI statevector is increased by two (to make room for RTHO and TEMPO). In the section for the BJT matrix setup: NXTRM=NXTRM+19> NXTRM=NXTRM+21 Subroutine MODCHK.f This routine preprocesses device model parameters and prints out a device summary. Add the eight new model variables, their names, and their default values. DIMENSION AMPAR(115) + AMPAR(123) DEFVAL(115) DEFVAL(123) IFMT(115)> IFMT(123) IVCHK(115)> IVCHK(123)
PAGE 111
105 IPAR holds the model parameter offset for each device type. Since the BJT is the second device in the list, the third and greater offsets must be increased by eight. DATA IPAR / 0, 1 4, 60, 72, 1 1 5 1 * IPAR / 0, 1 4, 68, 80, 1 23, / AMPAR / ... 7 6HXTI, 6HKF, 6HAF, 6HFC, 6HRTH, 6HCTH, 6HNEPI, 8 6HLTH, 6HWTH, 6HDTH, 6H0, 6H0, ... / DEFVAL holds the default values, and the new model parameters are set to zero. DEFVAL/... 3 ... 4 1.11 DO, 3.0D0, 0.0D0, 1 0D0, 0.5D0, 8*0.0D0 1 .../ IFMT holds the output format controls. IFMT/... 2 ... 2 1, 1, 2, 0, 0, 0, 0, 2, 1,1,2, 1, 1, 2, 2, 2, 4, 4, 4, 4, 4, 4, 0, 0, IVCHK defines whether the parameter can be negative. IVCHK/... 2 ... 2 0,0,0,0,0,0,0,0,0,0,0,0,0,1 ,0,0,0, 0, 0, 0, 0, 0, 0, 0, 0, 3 .../ SPICE requires several computed model parameters to be stored at the end of the model parameter list; the thermal parameters are inserted before these. Change all BJT modelvariable references (LOCV + x) where x is greater than 46 to (LOCV + x + 8), e. g. VALUE(LOCV+50)=FC*PC VALUE(LOCV+58)=FC*PC
PAGE 112
106 Subroutine READIN.f This routine processes element and device model definitions. Add the eight new model variables. DIMENSION AMPAR(115) > AMPAR(123) DATA IPAR / 0, 1 4, 60, 72, 1 1 5, / * IPAR / 0, 1 4, 68, 80, 1 23 / AMPAR/... 7 6HXTI, 6HKF, 6HAF, 6HFC, 6HRTH, 6HCTH, 6HLTH 8 6HWTH, 6HDTH, 6H0, 6H0, ... / Subroutine BJT.f This routine performs the initialization, matrix load, voltage limiting, and determines the operating point currents and derivatives for the dc and transient analyses. Add thermal resistance and temperature to the statevector. DIMENSION ... 2 ... CEXBC(1), GEQCBOO), TEMPO(1), RTHO(1) EQUIVALENCE ... 6 ... (GEQCBO(1 ),VALUE(1 9)), 7 0"EMPO(1 ), VALUE(20)), (RTHO(1 ), VALUE(21 )), DC model parameters section Additional needed variables. XJRB=VALUE(LOCM+1 7)* AREA C... THERMAL MODEL PARAMETERS PC=VALUE(LOCM+30) XTB=VALUE(LOCM+41) XTB=VALUE(LOCM+41) EG=VALUE(LOCM+42) XTI=VALUE(LOCM+43) RTH=VALUE(LOCM+47) ONEPI=VALUE(LOCM+49) OLTH=VALUE(LOCM+50) WTH=VALUE(LOCM+51) DTH=VALUE(LOCM+52)
PAGE 113
107 VTT=VT (so Vt can vary with each device's temperature) Change VALUE(LOCM + x) where x > 46 to VALUE(LOCM + x + 8), e. g. VCRIT=VALUE(LOCM+54) Â» VCRIT=VALUE(LOCM+62) Thermal model flag setup A multivalued flag, ITHERM, to direct control for different thermal models is created with the values shown in Table A4. Table A4 Key to thermal model flag: ITHERM 1 Thermal model using RTH. 2 Thermal model using predictive model (Nepi, L, W, D). 3 No thermal model. 4 Invalid region and thermal model using RTH. 5 Invalid region and predictive thermal mode. 6 Invalid region and no thermal model. Note that ITHERM is augmented by three when the current is nonpositive (CCHAT <= 0 or CC <= 0). The thermal model is not valid and should be avoided in these cases. After DC MODEL PARAMETERS and before INITIALIZATION insert: ITHERM=3 IF (RTH.NE.0.0D0) ITHERM=1 IF (ONEPI.NE.0.0D0) ITHERM=2 Device temperature, thermal resistance, and initial current The nominal temperature is referenced as VALUE(ITEMPS + ITEMNO) + CTOK in Kelvin. It is always the current
PAGE 114
108 ambient temperature whether it is the default temperature, .TEMP value, or TNOM value. In the DC MODEL PARAMETERS section: TNOM = VALUE(ITEMPS+ITEMNO)+CTOK At INITIALIZATION (INITF = 2) after line 25 insert: IF(ITHERM.EQ.3) GO TO 300 TEMPO(LXO+LOCT) = TNOM IF(ITHERM.EQ.I) RTHO(LX0+LOCT)=RTH IF(ITHERM.GTI) RTHO(LX0+LOCT)=0.0D0 This initializes the temperature to TNOM on the first iteration. THAT is the temperature predicted for the current iteration. Ideally this is computed from CCHAT, CBHAT, RTH, VBE, VBC, and TNOM, but CCHAT and CBHAT are not always valid. In the cases when the predicted currents are not available, the temperature at the end of the previous iteration is used for THAT. Near line 50 (INITF=4): IF(ITHERM.EQ.3) GO TO 21 0 THAT=TEMPO(LX0+LOCT) REGION=CCO(LX0+LOCT) IF(REGION.LE.O.ODO) ITHERM=ITHERM+3 GO TO 210 At end of COMPUTE NEW NONLINEAR BRANCH VOLTAGES section: REGION=CCHAT CBHAT= ... IF(REGION.LE.O.ODO) ITHERM=ITHERM+3 IF(ITHERM.LT3) THAT=RTHO(LX0+LOCT) *(DABS((VBEVBC)*CCHAT)+DABS(VBE'CBHAT))+TNOM
PAGE 115
109 Flow control To prevent the first convergence check when a thermal model is in use, the first line in BYPASS IF SOLUTION HAS NOT CHANGED section is changed to: IF (INITF.EQ.6.0R.ITHERM.NE.3.0R.n"HERM.NE.6) GO TO 200 This avoids premature convergence in dc sweeps. Update of temperaturedependent variables Temperature dependent model parameters IS, ISE, ISC, BF, BR, and VT are recomputed to reflect the local temperature. The following statements were added after the LIMIT NONLINEAR BRANCH VOLTAGES section: C. THERMAL MODEL TEMPERATURE UPDATE 21 0 IF(ITHERM.GT2) GO TO 300 RATIO=THAT/TNOM RATLOG=DLOG(RATIO) DTEMP=RATIO1.0D0 VTT=BOLTZ*THAT/CHARGE FACTLN=DTEMP*EGA/TT +XTI*RATLOG FACTOR=DEXP(FACTLN) CSAT=CSAT* FACTOR BFACTR=DEXP(XTB*RATLOG) BFM=BFM*BFACTR BRM=BRM*BFACTR C2=C2*DEXP(FACTLNA/ALUE(LOCM+7))/BFACTR C4=C4*DEXP(FACTLN/VALUE(LOCM+13))/BFACTR These equations are the same as those in the temperature update routine, TMPUPD.f.
PAGE 116
110 Thermal modifications to the conductances In the section following the CHECK CONVERGENCE section and before storage of the statevector variables line 750 is modified and the following code is inserted to correct the intrinsic device conductances for thermal effects: C C MODIFY CONDUCTANCES FOR THERMAL EFFECTS C 750 IF(ITHERM.GT3.AND.CC.GT0.0D0) ITHERM=ITHERM3 IF(ITHERM.GT2) GO TO 780 POWER = DABS(CC*(VBEVBC))+DABS(CB*VBE) TNEW = RTHO(LX0+LOCT)*POWER+TNOM GO TO (755,752), ITHERM If necessary RTH is predicted from VCB and the device's geometry. Here the temperature dependance of the thermal conductance, OKTH, has been added, C. PREDICT RTH 752 ARG1 =DSQRT(WTH*OLTH) RATIO=WTH/OLTH ARG2=PCVBC ARG2=DSQRT(2.0D6*EPSSIL*ARG2/(CHARGE*ONEPI))/ARG1 ARG3=DTH/ARG1 F1 =(0.058D0*ARG3+0.1 4D0)*ARG2+O.34D0*ARG3+O 28D0 F2=0.98D0+0.043D0*RATIO6.9D4*RATIO"2 0D0 1 +3.9D6*RATIO**3.0D0 OKTH=2.99D4/(TNEW99.0D0) RTHO(LX0+LOCT)=1 .ODO/(TWOPI*OKTH*2.0DO*ARG1 *F1 *F2) The equations for modifying the conductances apply to the commonemitter gparameters, so the hybridn parameters must be converted to gparameters, modified, and converted back to hybridn parameters. C. CON VERT HYBRIDPI TO GPARAMETERS 755 G11=GPI4GMU G12 = GMU G21 = GMGMU G22 = GO+GMU
PAGE 117
Ill TNEW is the temperature computed from the actual currents, not the linearly predicted ones, TCB is the temperature coefficient of base current with VBE held constant, and TCC is the temperature coefficient of collector current with VBE held constant. C. COMPUTE TEMPERATURE COEF. TIBE1 = (XTI+(EGVBE/VALUE(LOCM+3))/VTT) ARG1 = BFM*CBEN/VALUE(LOCM+7) ARG2 = ARG1*(XTIVALUE(LOCM+7)*XTBVBE/VTT) TCB = ((TIBE1XTB)*CBEARG2)/(TNEW*BFM*CB) TCC = TIBE1*(1 ODOCC*OIK*Q1/DSQRT(1.0D044.0DO*CBE*OIK))/TNEW Since the same equations hold in the ac analysis for the complex Yparameters, the unmodified values are saved. These electricalonly values will be corrected in the ac matrix load routine, ACLOAD.f. C..SAVE FOR AC ANALYSIS VALUE(LOCV+4)=GPI VALUE(LOCV+5)=GMU VALUE(LOCV+Â€)=GM VALUE(LOCV+7)=GO VALUE(LOCVfS)=TCB VALUE(LOCV+9)=TCC Next the gparameters are modified and converted back to hybridTi form. CAD JUST FOR THERMAL EFFECTS ARGC = TCC * RTHO(LX0+LOCT) ARGB = TCB * RTHO(LX0+LOCT) DENC = 1 .0D0 ARGC * POWER DENB = 1 .0D0 ARGB * POWER GNUM1 1 = ARGB * CB * CB GNUM21 = CB * CC GNUM12 = ARGB *GNUM21 GNUM21 = ARGC *GNUM21 GNUM22 = ARGC * CC * CC G1 1 =(G1 1 +GNUM1 1 )/DENB G1 2=(G1 2+GNUM1 2)/DENB G21 =(G21 +GNUM21 )/DENC G22=(G22+GNUM22)/DENC C. CON VERT BACK TO HYBRIDPI PARAMETERS 7 GPI = G11 +G12 GMU = G12
PAGE 118
112 GM = G21 G12 G0 = G22+G12 Changes for AC analysis Changes to the ac analysis simulation code are similar to the corrections to the conductances in the dc analysis. The only difference is the conductances become complex admittances and the thermal resistance becomes a complex thermal impedance, ZTH. The thermal impedance is computed from the singlepole model if CTH is specified on the model card. Otherwise ZTH is computed from RTH using the point heatsource model. Routine ACLOAD.f The following code adds two new temperature dependancies: OKTH is the thermal conductivity of silicon and OKAPPA is the thermal diffusivity of silicon. Simple relations were used to approximate the measured data for these variables [May67]. In the BJT section, check for a nonzero RTH in the statevector and for positive collector current. GEPR=VALUE(LOCM+1 9)* AREA C. THERMAL MODEL CC=VALUE(LOCT+2) RTH=VALUE(LOCT+20) IF(RTH.EQ.O.ODO.OR.CC.LE.O.ODO) GO TO 262 CB=VALUE(LOCT+3) POWER=DABS(CC*(VALUE(LOCT)VALUE(LOCT+1))) +DABS(CB*VALUE(LOCT)) TEMP= VALUE(LOCT+1 9) GPI=VALUE(LOCV+4) GMU=VALUE(LOCV+5) GM=VALUE(L0CV+6) GO=VALUE(LOCV+7) TCB=VALUE(LOCV+8) TCC=VALUE(LOCV+9) GO TO 263
PAGE 119
113 If no thermal model, 262 GPI=... Excess phase and imaginary parts of admittances, 263 XGM=... 270 GX=... Modify admittances C..THERMAL MODIFICATIONS TO ADMITTANCES IF(RTH.EQ.O.ODO.OR.CCLE.O.ODO) GO TO 280 C. CONVERT TO COMMONEMITTER YPARAMETERS Y1 1 =CMPLX(GPIfGMU,XCMCB+XCPI+XCMU) Y1 2=CMPLX(GMU,XCMCB+XCMU) Y21 =CMPLX(GMGMU,XGMXCMU) Y22=CMPLX(GO+GMU,XCMU) Singlepole model, IF (VALUE(LOCM+48).EQ.0.0D0) GO TO 273 ZTH=CMPLX(RTH,VALUE(LOCM+48)) GO TO 275 Computed from RTH using the temperature dependance of the thermal conductance and diffusivity, OKAPPA [Nis80]. C.COMPUTE ZTH FROM RTH 273 OKTH=2.99D2/(TEMP99.0D0) OKAPPA=OKTH*(TEMP99.0D0)/(2.33D0*(TEMP159.0D0)) ARG1 =DSQRT(OMEGA/(2*OKAPPA)) REFF=1 .0D0/(2.0D0*TWOPrOKTH*RTH) PHS=REFF*ARG1 OMAG=RTH*DEXP(PHS) ZTH=CMPLX(OMAG*DCOS(PHS),OMAG*DSIN(PHS)) Correct electricalonly admittances for selfheating, C. MODIFY 275 DENB=1 0D0/(1 .0D0TCB*POWER*ZTH) DENC=1 0D0/(1 0D0TCC*POWER*ZTH) ARG1=CC*CB Y1 1 =(Y1 1 +TCB*ZTH*CB*CB)*DENB Y1 2=(Y1 2+TCB*ZTH*ARG1 )*DENB Y21 =(Y21 +TCC*ZTH*ARG1 )*DENC Y22=(Y22+TCC*ZTH*CC*CC)*DENC C. CON VERT BACK TO HYBRIDPI GPI=REAL(Y1 1 )+REAL(Y1 2) GMU=REAL(Y1 2) GM=REAL(Y21 )REAL(Y1 2)
PAGE 120
114 GO= RE AL( Y22)+RE AL(Y1 2) XGM=AIMAG(Y21 )+AIMAG(Y22) XCPI=AIMAG(Y1 1 )AIMAG(Y1 2) XCMU=AIMAG(Y22) XCMBC=AIMAG(Y1 2)AIMAG(Y22) Stuff into matrix as usual, 280 LOCY=LYNL+NODPLC(LOC+24) ...
PAGE 121
APPENDIX B SPICE 2G.6 CODE CHANGES FOR MODELING OF SELFHEATING USING AN EXPLICIT TEMPERATURE NODE The following is a summary of the changes made to SPICE 2.G6 to implement selfheating in the BJT model. The local temperature of each device is computed for each iteration as a function of the power dissipation in the collectorbase spacecharge region (SCR). In the approach used here, temperature is represented explicitly by the addition of a new node to the device model, and hence a new row and column are needed in the device matrix equations. The dc, ac and transient analyses are all implemented using this thermal node approach. Data structure expansion The SPICE data structure must be expanded to include the new state variables and matrix locations for the thermal model. Five variables were added to the state vector shown in Table Bl Statevariables. VTMPO Device local temperature change. GTCO Collector thermal conductance. GTBO Base thermal conductance. CPO Device instantaneous power. 115
PAGE 122
116 Note: if more than a singlepole model is desired to model the thermal response then an additional energy and energycurrent statevariable will be needed for each additional pole. Subroutine Modifications What follows is a quick description of the changes needed to each routine to implement the thermal model. First the data structure must be expanded to include the parameters described above. This includes eight matrix locations and four state variables. ADDELT This routine adds an element and defines the size of linkedlist arrays for computation of offsets. Add eightelements to the matrix and one variable to the device. DATA LNOD(12) = 38 LNOD(12) = 46 LVAL(12) 4 * LVAL(12) = 5 NNODS(12) = 4 + NNODS(12) = 5 ALTER This routine changes the elements or device parameters for multiple SPICE runs. Add eight elements to the matrix and one variable to the device. DATA LNOD(12) = 38 LNOD(12) = 46 LVAL(1 2) = 4 + LVAL(12) = 5 ELPRNT In statement 300 change: NODPLC(LOC+36) Â» NODPLC(LOC+44)
PAGE 123
117 ERRCHK Add an external node to the BJT device line: DATA NNODS(12) 4~+ NNODS(12) = 5 FIND This routine presets storage for each input element. Add eight elements to the matrix and one variable to the device. DATA LNOD(12) = 38 * LNOD(12) = 46 LVAL(12) 4~* LVAL(12) = 5 MATLOC In statement 830: NODPLC(LOC+36) ~+ N0DPLC(L0C+44) Then add new elements to matrix: NODET=NODPLC(LOC436) NODPLC(LOC+37)=INDXX(NODE4,NODET) NODPLC(LOC+38)=INDXX(NODE5,NODET) NODPLC(LOC+39)=INDXX(NODE6,NODET) NODPLC(LOC440)=INDXX(NODET,NODE4) NODPLC(LOC+41 )=INDXX(NODET,NODE5) NODPLC(LOC442)=INDXX(NODET,NODE6) NODPLC(LOC+43)=INDXX(NODET,NODET) NODET is the external thermal node. MATPTR This routine builds the sparse matrix and it is here that the new nodes are added. In statement 530 change: NODPLC(LOC+36) Â» NODPLC(LOC+44)
PAGE 124
118 Then add the new nodes and elements to the matrix: NODET=NODPLC(LOC+36) CALL RESERV(NODE4,NODET) CALL RESERV(NODE5,NODET) CALL RESERV(NODE6,NODET) CALL RESERV(NODET,NODE4) CALL RESERV(NODET,NODE5) CALL RESERV(NODET,NODE6) CALL RESERV(NODET.NODET) The offset for the LXI statevector is increased by 4 (to make room for state variables). In the section for the BJT matrix setup: NXTRM=NXTRM+1 9 Â— NXTRM=NXTRM+23 MODCHK This routine preprocesses device model parameters and prints out a device summary. In statement 710 change: NODPLC(LOC+36) Â» N0DPLC(L0C+44) then move the reference to the new external node: NOD PLC(LOC+36)=NODPLC(LOC+6) READIN This routine processes element and device model definitions. DATA L2NOD(12) =36 * L2NOD(12) =44 To read the external thermal node, above statement 306 change: GO TO 308 GO TO 307 then after 306 add: 307 IF(NODPLC(ICODE+6).NE.0) GO TO 308 IFLD=7 IF (NODPLC(LOC+L2NOD(ID)).NE.O) GO TO 309 NODPLC(LOCfÂ€)=VALUE(IFIELD+6) GO TO 308 309 NODPLC(LOC+Â€)=0 308 CONTINUE (no change)
PAGE 125
119 The thermal node is temporarily placed in LOC+6 for reference in TOPCHK and is moved to LOC+36 in the MODCHK routine TOPCHK Add external thermal node reference: DATA NNODS(12) = 5 TRUNC In statement 80 change: NODPLC(LOC+36) Â» N0DPLC(L0C444) Changes for DC analysis The following are changes needed for operatingpoint and dcsweep analysis. Here we have created a mirror BJT routine that is executed in the event of a thermal analysis. This eliminates all of the "if thermal then do this" statements that would otherwise be needed in the original BJT routine. A drawback is the extra maintenance of a new model routine. BJT This routine performs the initialization, matrix load, voltage limiting, and determines the operating point currents and derivatives for the dc and transient analyses. In the case of a thermal transistor branch to thermal version after LOCM assignment: IF (VALUE(LOC+36).NE.1 ) THEN CALL BJTT(LOC) LOC=NODPLC(LOC) GO TO 10 ENDIF
PAGE 126
120 BJTT This routine performs the initialization, matrix load, voltage limiting, and determines the operatingpoint currents and derivatives for the dc and transient analyses for thermal transistors. Add thermal resistance, temperature, thermal conductance, and power current to the statevector. DIMENSION ... 3 VTMPO(1 ), GTCO(1 ), GTBO(1 ), CPO(1 ) EQUIVALENCE ... 7 (VTMPO(1 ),VALUE(20)). (GTCO(1 ),VALUE(21 )), (GTBO(1 ),VALUE(22)), 8 (CPO(1),VALUE(23)) then add the new node: NODET=NODPLC(LOC+36) Set the power to zero: CP=0.0D0 In the DC model parameters section some additional variables need to be denned: PC=VALUE(LOCM+30) XTB=VALUE(LOCM+41) EG=VALUE(LOCMf42) XTI=VALUE(LOCM+43) VTT=VT (so Vt can vary with each device's temperature) TNOM=VALUE(ITEMPS+ITEMNO)+CTOK Note: The nominal temperature is referenced as VALUE(ITEMPS+ITEMNO)+CTOK in Kelvin. It is always the current ambient temperature whether it is the default temperature, .TEMP value, or TNOM value.
PAGE 127
121 Initialize temperature flag (used to indicate wether device is in the proper region of operation for thermal iteration): NOTMP=. FALSE. Initialization section After statement 25 (INITF = 2) insert: VTMP=0.0D0 This initializes the temperature to the ambient. If the device is off there is no power dissipated. After statement 40 (INITF = 3) insert: VTMP=0.0D0 After statement 50 (INITF=4): VTMP=VTMPO(LX0+LOCT) IF (VBC.GT.0.0D0) NOTMP=.TRUE. After statement 60 (INITF=5): VTMP= VTMPO(LX1 +LOCT) IF (VBC.GT.0.0D0) NOTMP=.TRUE. After statement 70 (INITF=6): IF (VBC.GT.0.0D0) NOTMP=.TRUE. VTMPO(LX0+LOCT)=vTMPO(LX1 +LOCT) vTMP=(1.0D0+XFACT)*VTMPO(LX1+LOCT) XFACT*VTMPO(LX2+LOCT) GTCO(LX0+LOCT)=GTCO(LX1 +LOCT) GTBO(LX0+LOCT)=GTBO(LX1 +LOCT) After statement 100 (INITF=1), normal iteration: IF (VBC.GT.0.0D0) NOTMP=. TRUE. VTMP=VALUE(LVNIM1 +NODET)) and after statement 110: DELVTMP=VTMPVTMPO(LX0+LOCT) CCHAT=CCO(LX0+LOCT)+(GMO(LX0+LOCT)+GOO(LX0+LOCT))*DELVBE
PAGE 128
122 1 (GOO(LX0+LOCT)+GMUO(LX0+LOCT))*DELVBC4GTCO(LX0+LOCT)*DELVTMP CBHAT=CBO{LX0+LOCT)+GPIO(LX0+LOCT)*DELVBE4GMUO(LX0+LOCT)*DELVBC 1 4GTBO(LX0+LOCT)*DELVTMP If not forward active, set power to zero: IF (NOTMP) THEN CPHAT=0.0D0 ELSE CPHAT=CPO(LX0+LOCT)+DABS(CCO(LX0+LOCT)*DELVCE) 1 +DABS(CBO(LX0+LOCT)*DELVBE) ENDIF Add a convergence check for the temperature in the last position of the BYPASS IF SOLUTION HAS NOT CHANGED section: TOL=RELTOL*DMAX1 (DABS(VTMP),DABS{VTMPO(LX0+LOCT)))+VNTOL*1 D2 IF (DABS(DELVTMP).GE.TOL) GO TO 200 This avoids premature convergence in dc sweeps before the temperature has settled. After this add the lines: VTMP=VTMPO(LX0+LOCT) GTC=GTCO(LX0+LOCT) GTB=GTBO(LX0+LOCT) CP=CPO(LX0+LOCT) Update of temperaturedependent variables Temperaturedependent model parameters IS, ISE, ISC, BF, BR, EG, and VT are recomputed to reflect the local temperature increase. The following statements were added after the LIMIT NONLINEAR BRANCH VOLTAGES section: 210 CONTINUE IF (VTMP.LTO.ODO) THEN VTMP=0.0D0 NOTMP=.TRUE. ELSE LOOPTMP=.TRUE. DELTMP=1 D1 ODELTMP=1D1 VTMP=VTMP+DELTMP ENDIF 220 CONTINUE
PAGE 129
123 DTEMP=VTMP/TNOM RATI0=DTEMP+1 .ODO RATLOG=DLOG(RATIO) TLOC=TNOM+VTMP VTT=BOLTZ*(TLOC)/CHARGE EG=EGO(EGAP*TLOC*TLOC)/(TLOC+TGAP) FACTLN=DTEMP*EGA/TT+XTI*RATLOG FACTOR=DEXP(FACTLN) CSAT=VALUE(LOCM+1 )*AREA*FACTOR BFACTR=DEXP(XTB*RATLOG) BFM=VALUE(LOCM+2)*BFACTR BRM=VALUE(LOCM+8)*BFACTR C2=VALUE(LOCM+Â€)*AREA*DEXP(FACTLhWALUE(LOCM+7))/BFACTR C4=VALUE(LOCM+12)*AREA*DEXP(FACTLNA/ALUE(LOCM+13))/BFACTR These equations are equivalent to those in the temperature update routine, TMPUPD. Note: VTT is the new thermal voltage, and VT should be replaced by it throughout this subroutine. Computation of the new conductance and current terms The thermal model is valid only in the forwardactive region, so should only be computed in this case. If the device is not in the forwardactive region then temperature coefficients will be zero. The thermal transconductance is computed in the DETERMINE DC INCREMENTAL CONDUCTANCES section before statement 440: IF (NOTMP) THEN CP=0.0D0 GTB=0.0D0 GTC=0.0D0 ELSE IF (LOOPTMP) THEN CPHI=DABS(VCE*CC)+DABS(VBE*CB) CCHI=CC CBHI=CB CC=0.0D0 LOOPTMP=. FALSE. VTMP=VTMPDELTMP GOTO 220 ELSE CP=DABS(VCE*CC)+DABS(VBE*CB) GTC=(CCHICC)*ODELTMP GTB=(CBHICB)*ODELTMP
PAGE 130
124 ENDIF Check convergence After statement 700, add a convergence check for power: TOL=RELTOL*1 .0D3*DMAX1 (DABS(CPHAT),DABS(CP))+ABSTOL IF (DABS(CPHATCP).LE.TOL) GO TO 780 Load the matrix Save the state vector information: VTMPO(LX0+LOCT)=VTMP GTCO(LX0+LOCT)=GTC GTBO(LX0+LOCT)=GTB CPO(LX0+LOCT)=CP Compute and store the currents: 900 CEQBE=TYPE*(CC+CBVBE*(GM4GOfGPI)+VBC*(GOGEQCB) 1 VTMP*(GTC+GTB)) CEQBC=TYPE*(CC+VBE*(GM4GO)VBC*(GMU4GO)+VTMP*GTC) ... (to the end of section) IF(.NOT.NOTMP) THEN VALUE(LVN+NODET)=VALUE(LVN+NODET)+{CPVCE*(VCE*GO+VBE*GM 1 +vTMP*GTC*TYPE)VBE*(VBE*GPI+VTMP*GTB*TYPE)) ENDIF Stuff the conductance matrix: 950 ... (to the end of normal load) LOCY=LVN+NODPLC(LOC+37) VALUE(LOCY)=VALUE(LOCY)+GTC*TYPE LOCY=LVN+NODPLC(LOC+38) VALUE(L0CY)=VALUE(L0CY)4GTB*TYPE LOCY=LVN+NODPLC(LOC+39) VALUE(LOCY)=VALUE(LOCY) (GTC+GTBJTYPE LOCY=LVN+NODPLC(LOC+40) VALUE(LOCY)=VALUE(LOCY)VCE*GO*TYPE LOCY=LVN+NODPLC(LOC+41 ) VALUE(LOCY)=VALUE(LOCY)(VCE*GM+VBE*GPI)*TYPE LOCY=LVN+NODPLC(LOC+42) VALUE(LOCY)=VALUE(LOCY)+(VCE*(GM4GO)+VBE*GPI)*TYPE LOCY=LVN+NODPLC(LOC+43) VALUE(LOCY)=VALUE(LOCY)(GTC*VCE4GTB*VBE)*TYPE 1000 RETURN END
PAGE 131
125 Output of operatingpoint information DCOP This routine prints out operatingpoint information. The following changes allow thermal resistance, local temperature increase, and thermal conductance to be printed in response to the .OP command. DIMENSION 2 CCS(12), CBX(12), RX(12),DTMP(12), GTC(12), GTB(12) DATA ADTMR GTMP / 5HDELT, 3HGTC, 4HGTB / Change BJT section: In statement 320 and in the statement after 360 change: NODPLC(LOC+36) Â— NODPLC(LOCf44) then add: BETADC(KNTR)=CC(KNTR)/DSIGN(DMAX1 (DABS(CB(KNTR)),1 .OD20), 1 CB(KNTR)) DTMP(KNTR)=VALUE(LOCT+1 9) GTC(KNTR)=VALUE(LOCT+20) GTB(KNTR)= VALUE(LOCT+21 ) IF (MODEDC.NE.1) GO TO 325 IF (VALUE(LOCT+19) .EQ. 0.0D0) GO TO 355 WRITE (IOFILE.AFMT2) ADTMP,(DTMP(I),I=1,KNTR) WRITE (IOFILE.AFMT2) AGTC,(GTC(I),I=1,KNTR) WRITE (IOFILE.AFMT2) AGTB,(GTB(I),I=1 ,KNTR) 355 IF (MODEDC.NE.1) GO TO 360 These thermal variables are printed only if a thermal model is used. Changes for AC analysis Changes to the ac analysis simulation code are similar to the corrections to the change in the dc analysis.
PAGE 132
126 AC LOAD In the B JT section, change the end of linkedlist check: 260 IF ((LOC.EQ.O).OR.(NODPLC(LOC+44).NE.O)) GO TO 300 After the other admittance terms have been loaded the following code is added for the thermal model: IF(NODPLC(LOC+36).NE.1) THEN GTC=VALUE(LOCT+20) GTB=VALUE(LOCT+21) VBE=VALUE(LOCT) VCE=VALUE(LOCT)VALUE(LOCT+1 ) CC=VALUE(LOCT+2) LOCY=LYNL+NODPLC(LOC+37) VALUE(LOCY)=VALUE(LOCY)4GTC LOCY=LYNL+NODPLC(LOC+38) VALUE(LOCY)=VALUE(LOCY)+GTB LOCY=LYNL+NODPLC(LOC+39) VALUE(LOCY)=VALUE(LOCY)(GTC4GTB) LOCY=LYNL+NODPLC(LOC+40) VALUE(LOCY)=VALUE(LOCY)VCE*(GO+GMU)CC L0CY=LYNL+N0DPLC(L0C+41 ) VALUE(LOCY)=VALUE(LOCY)VCE*(GMGMU)VBE*GPICB LOCY=LYNL+NODPLC(LOC+42) VALUE(LOCY)=VALUE(LOCY)+VCE*(GM+GO)+VBE*GPI+CC+CB LOCY=LYNL+NODPLC(LOC+43) VALUE(LOCY)=VALUE(LOCY)GTC*VCEGTB*VBE ENDIF Table B2 Offsets for BJT matrix elements. LOC c B E c b e s ^ext C 24 10 B 25 34 11 E 26 12 c 13 35 27 14 15 32 37 b 16 17 28 18 38 e 19 20 21 29 39 s 33 31 T A ext 40 41 42 43
PAGE 133
APPENDIX C SOISPICE CODE CHANGES FOR MODELING OF SELFHEATING USING THE EXPLICITE TEMPERATURE APPROACH Introduction A physical chargebased model for partiallydepleted (PD) SOI MOSFETs, with selfheating, nonlocal impactionization current, parasitic BJT, and other optional features which can be important in shortchannel devices has been implemented in SOISPICE4. Similar changes were made to the fullydepleted (FD) SOI MOSFET model routines. PD SOI MOSFET Model User Reference The user specifies the use of the selfheating model by including the RTH parameter in the device card as explained in Table Cl. ZXXXXX ND NGF NS NGB (NB) (BJT) MNAME L=exp W=exp [AD=exp] [AS=exp] [AB=exp] [NRD=exp] [NRS=exp] [NRB=exp] [RTH=exp] [CTH=exp] [off] [IC = Vds, Vgfs, Vgbs, Vbs] Table Cl PD SOI MOSFET Device Line Explanation Name Description ND Drain node NGF Front gate node NS Source node NGB Back gate node NB Optional body node BJT Parasitic BJT (on) flag RTH Thermal resistance CTH Thermal capacitance 127
PAGE 134
128 PD SOI MOSFET Model Implementation The following figure shows how current and voltage are used as duals to power dissipation and temperature. The power of the device is applied as a current that generates a voltage as it flows through the thermal impedance, RTH and CTH, that is equivalent to the change in temperature of the device. The increase in temperature is used to update the device model parameters each iteration. If the parameter RTH is not specified on the device line, the above thermal subcircuit is not generated, and the standard isothermal model is used. The changes in SPICE2G.6 files made to implement selfheating in the PD SOI MOSFET model are marked with "Cdtz" comments in the Fortran source code. The implementations required changes to 17 subroutines and the addition of one new routine, all described below. Subroutines ADDELT and ALTER In these routines the matrix size defined by LNOD(16) was increased by 1 2 for the new admittance matrix elements. Subroutine FIND This routine checks for repeated element names and allocates memory for unique elements. The matrix size defined by LNOD(16) was increased by 12 for the new admittance matrix elements. Also, the IFORCE argument value of two is added to bypass the repeated name check. This is needed to allocate memory for the thermal capacitance,
PAGE 135
129 since all thermal capacitors have the same name. This is explained in the MODCHK section. Subroutine READIN This routine decodes the input deck. First L2NOD(16) is increased by 12 for thermal elements of the admittance matrix. The device line parameters RTH and CTH were added to the list. Subroutine MODCHK This routine processes the model parameters. In the NFD section where the parasitic resistances are added, check for RTH, add the temperature node to the matrix, and add the CTH capacitor if needed. IF (VALUE(LOCV+25).NE.0.0D0) THEN NUMNOD=NUMNOD+1 NODPLC(LOC+47)=NUMNOD VALUE(LCX)V+25)=1.0D0/VALUE(LOCV+25) IF (VALUE(LOCV+26).NE.0.0D0) THEN CALL FIND(ACGATOR,2,LOCTH,2) LOCvTH=NODPLC(LOCTH+1 ) NODPLC(LOCTH+2)=NUMNOD NODPLC(LOCTH+3)=1 VALUE(LOCVTH+1 )=VALUE(LOCV+26) NODPLC(LOCTH+4)=1 ENDIF ELSE NODPLC(LOC+47)=1 ENDIF All the thermal capacitors are added to the normal capacitor list with the name GATOR; this name was chosen arbitrarily. FIND is forced to bypass the namechecking by the 2 in the call (see above).
PAGE 136
130 Subroutines MATPTR and MATLOC These routines reserve memory for the linked list. The thermal matrix elements and state variables are added as shown in the linked list provided at the end of this document. Subroutine NFDFET This subroutine, by calling NFDID and NFDEC, performs the initialization, matrixload operations, and voltageiterate limiting, and determine the operatingpoint currents and charges and their respective derivatives for the PD SOI MOSFET device. The following is a description of the thermal changes relating to selfheating. Make a new common block for thermal arguments (used for FD devices also): COMMON /TMPARG/ GMTMRGIITMRGIRTMP.GITTMP.GPWFiTMP.GPWRGF.GPWRDS, & GPWRGB.GPWRBY.GD.GS Extend dimension and equivalence for the thermal statevariables. Get temperature node (this will be a flag for temperature modeling). NTMP=NODPLC(LOC447) Variables VTMP (temperature change) and CPWR (power dissipation) are treated the same as any other voltages and currents. The temperature change is limited to positive values. Add thermal arguments to NFDID call for normal and inverse modes: CALL NFDID(VGFS,VDS,VGBS,VBS,LOC,QD,QGF,QS,QGB,QBY & & CBD,CBGB,CBS,CDGB,CSGB,CGBD,VTMP,CPWR,CPWRTMP & CGFTMPCDTMPCGBTMP.CBYTMPCSTMP)
PAGE 137
131 In the transient analysis, add thermal parameters to NFDEC call and include in tranconductance computations for normal and inverse modes: CALL NFDEC(VGFD,VGFS,VGFB,VGBD > VGBS,VGBB,XNQFF,XNQFB, & & CDTMP.CSTMP.CGFTMP.CGBTMP.CBYTMRGCDTMP.GCSTMRGCGFTMR & GCGBTMRGCBYTMP) GCGFS=(GCGFGF+GCGFD+GCGFB+GCGFTMP) GCDS=(GCDGF+GCDDfGCDB+GCDGB4GCDTMP) GCBS=(GCBGF+GCBD+GCBB+GCBGB+GCBYTMP) GCGBS=(GCGBB+GCGBD+GCGBGB+GCGBTMP) GCSS=(GCSGF+GCSDtGCSB+GCSGB+GCSTMP) Conditionally compute currentvector element for power and add temperature parameters to other vector elements for normal and inverse modes, and load power element of current vector. Also check if drain and source resistances need updating: IF (NTMP.NE.1) THEN VALUE(LVN+NTMP)=VALUE(LVN+NTMP)+CPWREQ IF (VTMPGT.O.ODO) THEN GDPFM3D GSPR=GS ENDIF ENDIF Load thermal elements of the admittance matrix. Subroutine NFDID This subroutine, by calling NFDMOD, numerically calculates the derivatives of the operatingpoint currents and charges for both the forward and inverse modes of operation of the NFD/SOI device. The same changes to the common block and subroutine arguments done for NFDFET are repeated here. An additional call to NFDMOD is made to find the temperature derivatives.
PAGE 138
132 Subroutine NFDEC This subroutine calculates the transcapacitance terms used in transient and AC analysis from the intrinsic and extrinsic components of charge in the NFD/SOI device. Thermal transient conductance terms are also computed in this routine. Subroutine NFDMOD This subroutine, the primary model routine, calculates the channel transport, impactionization, junction leakage, and if requested the parasitic bipolar transistor currents of the PD SOI MOSFET for the given internal device voltages, as well as the intrinsic charge components that define the charging currents. After the model parameters are loaded and before any computations are made, SOITMP is called to update the temperature dependent parameters for the local temperature. The local thermal voltage VTT is substituted everywhere the global thermal voltage VT was used. It is important to assign the model parameter values to local variables before the SOITMP call and not to directly access the model parameter values (e.g. VALUE(LOCM+?)) in subsequent calculations. Otherwise these model parameters would not reflect changes in local temperature, which could cause inaccurate results.
PAGE 139
133 Subroutine SOITMP This subroutine updates all of the temperature dependent parameters for each device on each iteration. It is called by the NFDMOD routine. The parameters affected are shown in Table C2. Table C2 Temperature dependent parameters. NAME Description XNIN Intrinsic carrier density EGFET Silicon energy gap VTT Local Thermal Voltage YMT TT rÂ»C2 T LDD carrier mobility YMT TRTT Minority earner mobility in highdoped body uo Zero field electron mobility VSAT Carrier saturated drift velocity WKF Frontgate work function difference WKB Backgate work function difference VFBF Frontgate flatband voltage VFBB Backgate flatband voltage JRO Bodysource junction recombination current coefficient ALPHA Impactionization parameter a 0 TAUG Carrier lifetime in lightly doped region VBI Source to bulk builtin potential PHIB Twice fermi potential of body RLDD LDD sheet resistivity BETA b FSAT fsat GD Drain parasitic resistance GS Source parasitic resistance These are the same dependencies used in TMPUPD implemented by Glen Workman [Wor96]. If temperature dependencies are changed or added to TMPUPD, corresponding changes must also be made to SOITMP.
PAGE 140
134 Subroutine DCOP This routine generates the output for the operating point. If selfheating is modeled for a device, the change in temperature, power dissipated, and relevant conductance and capacitance terms are printed. Subroutine ACLOAD This routine loads the complex admittance matrix for the ac analysis. If there is a device with selfheating, the thermal elements of the matrix are loaded. See description of NFDFET routine. Data Structures The data structures for the PD SOI MOSFET device (ID=16) and model (ID=26), and for the print (ID=41) and plot (ID=42) are given in Table C3 for SOISPICE4.3i. All other linked lists remain the same as in the original SPICE2G.6. Table C3 Offsets for PD SOI MOSFET matrix elements. LOC D Gf S Gb B d s b T D 38 13 Gf 39 14 15 16 49 S 40 17 Gb 41 18 19 20 50 B 42 21 d 22 23 24 43 25 26 51 s 27 28 29 30 44 31 52 b 32 33 34 35 36 45 53 T 54 55 56 57 58 48
PAGE 141
135 PD SOI MOSFET Transistor (ID=16) U 1 : subckt info i +nn : next pointer LOCV +00: ~U 1 Â• i nrv LUvV +nr T UZ 1 lvJ +n9: ngf +ni+Uo. +(\A 'U4 ne lib +(14+U5 nyu +n^+UO. +UO nh riQ +nA+fl7 + U/ nap +n7+UO nsp +nA+UO. nbp +no+ 1 n 1 1 icxjtri poimer + in+ IU. 411 UIT + i i Â• + 19 + 1 z + 1941 7 T 1 O tna,nap; + 1 * + lo. + 1/1 + 1 /!Â• + 1C T IO + Tv + 1 A T 1 0 t'nof nhn^ viyMiup/ + 1 AT 1 0. + 17 V 1 'b , I IdfJJ + 17+ lft ' 1 o wiyujiup; + 1 ft+ 10 f'nnh ncr^ Â• ingu.nsp/ ^ i y. +9n uigD.nDp; +9n+ZU. +91 vnD.nDp; +91 Â• +z l . +99 map,na; +ZZ. +91 +ZO . (nap.ngf) +23: +Z4 (nap.ngo) +z4: +Z5 inap.nsp; i Of . +OA +ZO (nap.nDp; +ZO. +97 uisp.nyT; +97+9A vnsp,nsj +9A+ZO. +90 insp.ngo; +90+ZT. inspjiup; +^1 +^9 fnKn norf\ inup,ngi; +13 /nhn n/"irÂ™^ Uiup,nyu; +\4 T o4 (nup.nD; +4o: +m Wiup,n COVLGBB RTH CTH nt (nt.nf) (ngf.nt) (ngb.nt) (ndp.nt) (nsp.nt) (nbp.nt) (nt.ngO (nt.ngb) (nt.ndp) (nt.nsp) (nt.nbp)
PAGE 142
136 PD SOI MOSFET Model (ID=26) 01: subcktinfo +00: next pointer LOCM +00: model name +01: LOCM +01: VFBF +02: model tvDe +02: VFBB +03: TOXF +04: TOXB +05: WKF +06: WKB +07: NQFF +08: NQFB +09: NSUB + 10: NGATE + 11: FNK +12: FNA +13: TF +14: TB +15: THALO + 16: NBL + 17: NBH + 18: NHALO + 19: UO +20: THETA +21: VSAT +22: TPG +23: TPS +24: NLDS +25: NDS +26: LLDS +27: HDD +28: LDIFF +29: ZETA +30: ETA +31: LMOD +32: ALPHA +33: BETA +34: BFACT +35CGFDO +36CGFSO +37: CGFBO +38: RD +39: RS +40: RB +46: JRO +41: TAUO +47: M +42: DL +48: VJMIN +43: DW +49: unavailable +44: RHOSD +50: FVBJT +45: RHOB State Variable Table (PD SOI MOSFET Model) Lxi+00: vbso(VBS) +01: vbdo(VBD) +02: vgfso(VGFS)
PAGE 143
137 +03 : vgfdo(VGFD) +04 : vgbso (VGBS) +05 : vdso (VDS) +06 : cicho(ICH) +07 : ciro OR) +08 : cigto (IGT) +09 : cigioOGI) + 10 : cbjto(IBJT) + 11 : gidgfso (GDGF) + 12 : giddso(GDO) + 13 : gidbso (GDB) + 14 : giigfso (GGIGF) + 15 : giidso (GGID) + 16 : giibso (GGIB) + 17 : girbso GRB) + 18 : qdo(QD) + 19 : cqdo (IQD) +20 : qgfo (QGF) +21 : cqgfo (IQGF) +22 : qso (OS) +23 : cqso (IQS) +24 : qgbo (QGB) +25 : cqgbo (IQGB) +26 qbo(QB) +49 : vtwo +27 : cqbo (IQB) +50 : vdsato +28 cdgfo +51 : vdsato +29 cddo +52 : leffo +30 cdbo +53 cgbdo +31 cdso +54 cdgbo +32 cgfgfo +55 csgbo +33 cgfdo +56 qno +34 cgfbo +57 vtmpo (temperature) +35 cgfso +58 cpwro (power) +36 csgfo +59 gmtmpo +37 csdo +60 giitmpo +38 csbo +61 girtmpo +39 csso +62 cgttmpo +40 cgbgbo +63 cdtmpo +41 cgbbo +64 cstmpo +42 cgbso +65 cgbtmpo +43 cbgfo +66 cbytmpo +44 cbdo +67 gpwrtmpo +45 cbgbo +68 gpwrgfo +46 cbbo +69 gpwrdso +47 cbso +70 gpwrgbo +48. vtso +71 gpwrbyo Comments (PD SOI MOSFET Model): "Model Type" is +1 for NNFDSOI and 1 for PNFDSOI. The following preprocessing is performed in subroutine MODCHK: Location: Stored Term:
PAGE 144
138 LOCM+NZPARAM+l LOCM+NZPARAM+2 LOCM+NZPARAM+3 LOCM+NZPARAM+4 LOCM+NZPARAM+5 LOCM+NZPARAM+6 LOCM+NZPARAM+7 LOCM+NZPARAM+8 LOCM+NZPARAM+9 LOCM+NZPARAM+10 LOCM+NZPARAM+l 1 LOCM+NZPARAM+l 2 LOCM+NZPARAM+l 3 LOCM+NZPARAM+14 LOCM+NZPARAM+l 5 LOCM+NZPARAM+l 6 LOCM+NZPARAM+l 7 LOCM+NZPARAM+l 8 LOCM+NZPARAM+l 9 VBI (SB builtin potential) CB=EPSSIL/TB COXF=EPSOX/TOXF COXB=EPSOX/TOXB QB=CHARGE*TB*NBL PHIB=2"VTDLOG(NBL/NI) TLDD=TFTHALO XALPHA=l+CB/COXF XLC=TB( CB/(2*COXF*XALPHA)) l/2 XMUBH (minoritycarrier mobility in highdoped body) (unused> ESLOPE=CHARGE*NLDS/EPSSIL TAUG=2*TAUO/(l +NBL/NO) XMOBFAC1 =THETA*COXF/(2*EPSSIL) XBFACT=BFACT"XMOBFACl / (2XALPHA) Note: NZPARAM=50 for current number of input parameters. Additional COMMON blocks: COMMON /NFDARG/ CICH.CIR.CIGT.CIGI.CBJT.GIDGFS.GIDDS.GIDBS, GIIGFS,GIIDS,GIIBS,GIRBS,VTS,VTW,XLEFF,VDSAT COMMON /TMPARG/ GMTMRGIITMRGIRTMRGITTMRGPWRTMP, GPWRGF.GPWRDS.GPWRGB.GPWRBY
PAGE 145
139 Output Variables (IP=41.42) 01 subckt info LOC+00 next pointer LOCV +00: Element Name +01 LOCV +01: Device Name +02 Device ID / LOC +02: Oper Var Name +03 KOFF (LOCT/LOCV offset) +04 ISEQ +05 type=9 +06 KLOC (KLOCT/KLOCV offset) +07 ISTATE (S=0, P=l) Comments: For ID=41,42 and type (NODPLC(LOC+5)) = 0, 1, and 28, the definitions of the variables remain as defined in the "Program Reference for SPICE2".
PAGE 146
APPENDIX D SPICE CIRCUIT DESCRIPTIONS USED FOR SIMULATION EXAMPLES Single BJT CommonEmitter Amplifier This is a single NPN commonemitter amplifier circuit with an opencircuit load used to generate Figure 41.. Dc operating point and ac analyses are executed. COMMONEMITTER AMPLIFIER CIRCUIT VBE 2 0 .82 AC 1 VCE 1 0 5 LT1 3 1G Q1 3 2 0 0 4H3X50N XZTH 4 0 ZTH *Threepole Model .SUBCKT ZTH 1 4 RTH1 1 2 92.4 CTH1 1 2 1 .02u RTH2 2 3 142.8 CTH2 2 3 0.187U RTH3 3 4 44.8 CTH3 3 4 55.6n .ENDS * .MODEL H3X50N NPN (IS=1 .875E16 XTI=3 EG=1.16 VAF=60 VAR=4.5 + BF=146.1 ISE=1.875E19 NE=1.4 IKF=7.5E2 XTB=2 BR=10 + ISC=1.875E14 NC=1.8 IKR=7.5E2 RC=29.4 CJC=2.818E13 + MJC=0.24 VJC=0.97 FC=0.5 CJE=3.9E13 MJE=0.51 VJE=0.872 + TR=4E9TF=17.85E12 ITF=1 .155 XTF=78.81 VTF=10 PTF=0 + XCJC=0.1756 CJS=1.689E13 VJS=0.75 MJS=0 RE=1.333 RB=35.18 + RBM=0 KF=0 AF=1 EGAP=7.02E4 TGAP=1 108) 'Simulation Section .OPTIONS NUMDGT=7 LIMPTS=2000 RELTOL=1 e5 ITL1=1000 ITL2=1000 .WIDTH OUT=80 .OP .AC DEC 5 1 1G .PRINT AC V(3) .END 140
PAGE 147
141 741 Operational Amplifier Circuit This is the 741 operational amplifier found in Grey and Meyer [Gra84]. The schematic in shown in Figure 42, and the simulation results and shown in Figures 43 and 44. Dc operating point, dc transfer and ac analyses are executed. SELFHEATING 741 OPAMP CIRCUIT VCC 1 0 DC 15 VEE6 0DC15 VI1 8 0DC0 VI2 9 0 DC .835M AC 1 R1 1561K R2 16 61K R3 14 6 50K R4 5 6 5K R5 2 3 39K R6 24 25 27 R7 25 26 22 R8 20 6100 R9 19 6 50K R10 22 23 40K CC1317.65P Q1 7 8 10 6 51 H3x50N 02 7 9 11 6 52 H3x50N Q3 12 4 101 53 H3x50P Q4 13 411 1 54H3x50P Q5 12 1415 6 55H3x50N O6 13 14 16 6 56 H3x50N Q7 1 12 14 6 57H3x50N Q8 7 7 1 1 58 H3x50P Q9 4 7 1 1 59 H3x50P Q104 3 5 660H3X50N Q1 1 3 3 6 6 61 H3x50N Q1222 1 1 62 H3x50P Q13A21 2 1 1 63 H3x50P .25 Q13B17 2 1 1 64H3x50P.75 Q141 21 24 6 65 H3x50N Q161 13 19 6 66 H3x50N Q17 17 19 20 6 67 H3x50N Q18 21 22 23 6 68 H3x50N Q1 9 21 21 22 6 69 H3x50N 020 6 23 26 1 70 H3x50P Q23A6 17 23 1 73 H3x50P XZTH1 51 0 ZTH XZTH2 52 0 ZTH XZTH3 53 0 ZTH XZTH4 54 0 ZTH
PAGE 148
142 741 opamp circuit continued. XZTH5 55 0 ZTH XZTH6 56 0 ZTH XZTH7 57 0 ZTH XZTH8 58 0 ZTH XZTH9 59 0 ZTH XZTH10 60 0 ZTH XZTH1 1 61 0 ZTH XZTH12 62 0 ZTH XZTH13 63 0 ZTH XZTH14 64 0 ZTH XZTH15 65 0 ZTH XZTH16 66 0 ZTH XZTH17 67 0 ZTH XZTH18 68 0 ZTH XZTH19 69 0 ZTH XZTH20 70 0 ZTH XZTH23 73 0 ZTH *Threepole Model .SUBCKT ZTH 1 4 RTH1 1 2 92.4 CTH1 1 2 1 .02u RTH2 2 3 142.8 CTH2 2 3 0.187U RTH3 3 4 44.8 CTH3 3 4 55.6n .ENDS * .MODEL H3X50N NPN (IS=1.875E16 XTI=3 EG=1.16 VAF=60 VAR=4.5 + BF=146.1 ISE=1.875E19 NE=1.4 IKF=7.5E2 XTB=2 BR=10 + ISC=1.875E14 NC=1.8 IKR=7.5E2 RC=29.4 CJC=2.818E13 + MJC=0.24 VJC=0.97 FC=0.5 CJE=3.9E13 MJE=0.51 VJE=0.872 + TR=4E9 TF=17.85E12 ITF=1.155 XTF=78.81 VTF=10 PTF=0 + XCJC=0.1756 CJS=1.689E13 VJS=0.75 MJS=0 RE=1.333 RB=35.18 + RBM=0 KF=0 AF=1 EGAP=7.02E4 TGAP=1108) * .MODEL H3X50P PNP (IS=1.02E16 XTI=3 EG=1.16 VAF=30 VAR=4.5 + BF=70.11 ISE=1.02E19 NE=1.4 IKF=7.5E2 XTB=2 BR=7 + ISC=1.02E14 NC=1.8 IKR=7.5E2 RC=38 CJC=4.27E13 + MJC=0.3 VJC=1 .23 FC=0.5 CJE=4.8E13 MJE=0.57 VJE=0.88 + TR=4E9 TF=33.91E12 ITF=0.7127 XTF=45.14 VTF=10 PTF=0 + XCJC=0.1756 CJS=1.689E13 VJS=0.75 MJS=0 RE=1.333 RB=37.4 + RBM=0 KF=0 AF=1 EGAP=7.02E4TGAP=1108) Â•Simulation Section .OPTIONS NUMDGT=7 TNOM=25 ACCT OP .DC VI2 0 3m 10u .AC DEC 5 1 100meg .PRINT DC V(25) .PRINT AC VM(25) END
PAGE 149
143 PinDriver Transient This is the pindriver circuit to demonstrate the transient effects of selfheating. The schematic for this circuit is shown in Figure 45, and the results are shown in Figures 46 and 47. Â•THERMAL PINDRIVER CIRCUIT VIN 3 2 5 pulse (2 18 1u 5n 5n 500u 1000u) Q1 2 341 10 H3X50P Q2 7761 11 H3X50P Q3 5 5 42 12 H3X50N Q41 5 62 13H3X50N 11 1 5 DC 1m 12 7 2 DC 10m VEE 2 0 DC 10 VCC1 0DC10 XZTH1 10 0ZTH XZTH2 11 OZTH XZTH3 12 0 ZTH XZTH4 13 0 ZTH Threepole Model SUBCKT ZTH 1 4 RTH1 1 2 92.4 CTH1 1 2 1.02u RTH2 2 3142.8 CTH2 2 3 0.187u RTH3 3 4 44.8 CTH3 3 4 55.6n .ENDS .MODEL H3X50N NPN (IS=1.875E16 XTI=3 EG=1.16 VAF=60 VAR=4.5 + BF=146.1 ISE=1.875E19 NE=1.4 IKF=7.5E2 XTB=2 BR=10 + ISC=1.875E14 NC=1.8 IKR=7.5E2 RC=29.4 CJC=2.818E13 + MJC=0.24 VJC=0.97 FC=0.5 CJE=3.9E13 MJE=0.51 VJE=0.872 + TR=4E9TF=17.85E12 ITF=1.155 XTF=78.81 VTF=10 PTF=0 + XCJC=0.1756 CJS=1.689E13 VJS=0.75 MJS=0 RE=0 RB=35.18 + RBM=0 KF=0 AF=1 EGAP=7.02E4 TGAP=1108) * MODEL H3X50P PNP (IS=1.02E16 XTI=3 EG=1.16 VAF=30 VAR=4.5 + BF=70.11 ISE=1.02E19 NE=1.4 IKF=7.5E2 XTB=2 BR=7 + ISC=1 02E1 4 NC=1 .8 IKR=7.5E2 RC=38 CJC=4.27E1 3 + MJC=0.3 VJC=1 .23 FC=0.5 CJE=4.8E13 MJE=0.57 VJE=0.88 + TR=4E9TF=33.91E12 ITF=0.7127 XTF=45.14 VTF=10 PTF=0 + XCJC=0.1756 CJS=1.689E13 VJS=0.75 MJS=0 RE=0 RB=37.4 + RBM=0 KF=0 AF=1 EGAP=7.02E4 TGAP=1 108) Â•Simulation Section OPTIONS LIMPTS=2000 RELTOL=1 E4 METHOD=GEAR .TRAN 1 U 250U .PRINT TRAN V(3) V(7) V(10) V(13) END
PAGE 150
144 Thermal Coupling of ThreeFinger Emitter This is the threefinger emitter circuit to demonstrate the thermal coupling effects that are now possible to simulate. The schematic for this circuit is shown in Figure 48, and the results are shown in Figure 49. THREE FINGER NPN UHF 3X50UM DC TEST IB 01 .15M AC 1 VCE 2 0 5 QNPN1 2 1 0 0 4 H3X50N QNPN2 2 1 0 0 8 H3X50N QNPN3 2 1 0 0 12 H3X50N RTH1 4 5 280 HTH12 5 6 VTH2 100 HTH13 6 7 VTH3 20 VTH1 7 0 0 RTH2 8 9 280 HTH23 910VTH1 100 HTH21 1011 VTH3100 VTH2 110 0 RTH3 12 13 280 HTH31 1314VTH1 20 HTH32 1415 VTH2 100 VTH3 15 0 0 * .MODEL H3X50N NPN (IS=1 .875E16 XTI=3 EG=1.16 VAF=60 VAR=4.5 + BF=146.1 ISE=1.875E19 NE=1.4 IKF=7.5E2 XTB=2 BR=10 + ISC=1.875E14 NC=1.8 IKR=7.5E2 RC=29.4 CJC=2.818E13 + MJC=0.24 VJC=0.97 FC=0.5 CJE=3.9E13 MJE=0.51 VJE=0.872 + TR=4E9TF=17.85E12 ITF=1.155 XTF=78.81 VTF=10 PTF=0 + XCJC=0.1756 CJS=1.689E13 VJS=0.75 MJS=0 RE=1.333 RB=35.18 + RBM=0 KF=0 AF=1 EGAP=7.02E4 TGAP=1 108) 'Simulation Section .OPTIONS NUMDGT=7 LIMPTS=2000 RELTOL=1 e5 ITL1=1000 ITI_2=1000 .WIDTH OLTT=80 .DC VCE .1 10 0.05 .PRINT DC &(QNPN1,IC) &(QNPN2,IC) .END
PAGE 151
145 SOI MOSFET Characteristic Curves This is a single nchannel partiallydepleted SOI MOSFET circuit for generating the characteristic curves with and without selfheating. The results for this circuit are shown in Figure 65. PD SOIMOSFET characteristic curves with selfheating *voltage source VCC 5 0 DC 2.5 VGB1 3 0 0 VGF1 11 0 DC 2 'current detection VDUM1 5 21 0 VDUM2 5 22 0 ZN1 21 11 0 3 BJT MODN L=0.2E6 W=10E6 AD=10P AS=10P RTH=K CTH=5P ZN2 22 11 0 3 BJT MODN L=0.2E6 W=10E6 AD=10P AS=10P * MODEL MODN NNFDSOI (insert model here) 'Simulation Section .DC VCC 0 3 0.1 VGF1 1 2 .5 .PRINT DC l(VDUM1) l(VDUM2) .OPTION LIMPTS=2000 METHOD=GEAR ITL1=500 ITL2=200 RELTOL=1 E5 .END
PAGE 152
146 SOI MOSFET Step Response This is a single nchannel partiallydepleted SOI MOSFET circuit for generating the transient stepresponse with and without selfheating. The results and schematic for this circuit are shown in Figure 66. PD SOIMOSFET step response 'Voltage source VDD5 0 2.5 VGF11 0 PULSE (2 310N10N10N50N120N) VGB 4 0 0 Â•first stage: ZN1 5 11 0 40 BJT MODN L=0.2E6 W=10E6 AD=10P AS=10P RTH=8K CTH=115P ZN2 5 11 0 40 BJT MODN L=0.2E6 W=10E6 AD=10P AS=10P * .MODEL MODN NNFDSOI (insert model here) "Simulation Section TRAN 10N 5U '.PRINT TRAN V(21) V(22) '.PRINT TRAN V(22) PRINT TRAN &(ZN1 ,TEMP) V(1 1) .OPTION LIMPTS=2000 METHOD=GEAR ITL1=500 ITL2=200 RELTOL=1 e4 .END
PAGE 153
147 SOI MOSFET Inverter Switching This is a complementary partiallydepleted SOI MOSFET inverter circuit for generating the transient switching waveforms with and without selfheating. The schematic for the circuit is shown in Figure 67, and the results are shown in Figure 68. PD SOI MOSFET Inverter circuit VON 1 0 PULSE 0 3 2N 5N 5N 50N 1 00N VDD903 VGB1 20 0 0 'Inverter * without RTH and CTH *ZN0 2 1 0 20 BJT MODN L=.2E6 W=1 1 6E6 AD7.54P AS=12.6P IC=0,0,0,0.9 *ZP0 2 1 9 20 BJT MODP L=. 2E6 W=15.7E6 AD=10.2P AS=23.6P IC=0,0,0,0.9 * with RTH and CTH ZNO 2 1 0 20 BJT MODN L=.2E6 W=11.6E6 AD=7.54P AS=12.6P + RTH=7.1 431 Ef03 CTH= 9.7547E1 1 IC=0,0,0,0.9 ZPO 2 1 9 20 BJT MODP L=.2E6 W=15.7E6 AD=10.2P AS=23.6P + RTH=5.71 34E+03 CTH= 1 .3202E1 0 10=0,0,0,0.9 CLOAD 2 0 5p * .MODEL MODN NNFDSOI (insert model here) .MODEL MODP PNFDSOI (insert model here) Â•Simulation Section TRAN 10N 2U .PRINT TRAN V(2) &(ZN0,TEMP) .OPTION ACCT NODE OPTS LIST LIMTIM=50 METHOD=GEAR + PIVTOL=1E30 + GMIN=1 00E1 6 RELTOL=1 00E03 ABSTOL=1 .00E1 2 + VNTOL=1 .00E06 TRTOL=7.00E+O0 + ITL1 =1 00E+O2 fTL2=1 00E+O2 ITL3=4.00E+00 ITL4=2.50EfO2 + ITL5=0.00E+O0 LVLTIM=2.OOE40O TNOM=+2.50E+01 LIMPTS=99999 + CHGTOL=1 .OOE20 .END
PAGE 154
BIBLIOGRAPHY [Bro93] J. S. Brodsky, D. T. Zweidinger, and R. M. Fox, "Physicsbased multiplepole models for BJT selfheating", Proc. IEEE Bipolar Circuits and Technology Meeting, pp. 249252, 1993. [Bro97a] J. S. Brodsky, R. M. Fox, D. T. Zweidinger, and S. Veeraraghavan, "Physicsbased dynamic selfheating model for SOI MOSFETs," to be published in IEEE Transactions on Electron Devices, vol. 44, no. 6, June 1997. [Bro97b] J. S. Brodsky, Phvsicsbased thermal impedance model for the simulation of selfheating in semiconductor devices and circuits. Ph.D. Dissertation, University of Florida, Gainesville, 1997. [Dav92] C. Davis, G. Bajor, J. Butler, T Crandell, J. Delgado, T. Jung, Y. KhajehNoori, B. Lomenick, V. Milam, H. Nicolay, S. Richmond and T. Rivoli, "UHF1 : A high speed complementary bipolar analog process on SOI", Proc. IEEE Bipolar Circuits and Technology Meeting, pp. 260261, 1992. [Dav88] W. F. Davis and M. L. Lidke, "The effect of thermal feedback within the bipolar transistor on GummelPoon model accuracy," Motorola Internal Report, May 27, 1988. [Den89] R. T. Dennison and K. M. Walter, "Local thermal effects in high performance bipolar circuits/devices," Proc. IEEE Bipolar Circuits and Technology Meeting, pp. 164167, 1989. [Ebe54] J. J. Ebers and J. L. Moll, "Largesignal behavior of junction transistors," Proc. IRE, vol. 42, pp. 17611772, Dec. 1954. [Fos96] J. G. Fossum, R. M. Fox and D. T. Zweidinger, "SOISPICE4 (Ver. 4.4) SiliconOnInsulator MOSFET SPICE with SelfHeating," Programmer's Reference Manual, University of Florida Department of Electrical and Computer Engineering, Oct. 1996. [Fox91a] R. M. Fox and S.G. Lee, "Predictive modeling of themal effects in BJTs," Proc. IEEE Bipolar Circuits and Technology Meeting, pp. 8992, Sept. 1991. 148
PAGE 155
149 [Fox91b] R. M. Fox and S.G. Lee, "Scalable smallsignal model for BJT selfheating," IEEE Electron Device Letters, vol. 12, no. 12, pp. 649651, Dec. 1991. [Fox91c] R. M. Fox and S.G. Lee, "Thermal parameter extraction for bipolar circuit modeling", Electronics Letters, vol. 27, no. 19, pp. 17191720, 1991. [Fox93a] R. M. Fox and J. S. Brodsky, "Effects of selfheatinginduced negative output conductance in SOI circuits," SOI Conf. '93, Oct. 1993, [Fox93b] R. M. Fox, S.G. Lee, and D. T. Zweidinger, "The effects of BJT selfheating on circuit behavior", IEEE J. SolidState Circuits, vol. 28, no. 6, pp. 678685, June 1993. [Fuk76] K. Fukahori and R R. Gray, "Computer simulation of integrated circuits in the presence of electrothermal interaction," IEEE J. SolidState Circuits, vol. SC11, pp. 834846, Dec. 1976. [Gan92] P. Ganci, et al., "Selfheating in highperformance bipolar transistors fabricated on SOI substrates," Technical Digest IEEE International Electron Devices Meeting, pp. 417420, 1992. [Gao89] G.B. Gao, M.Z. Wang, X. Gui, and H. Morkoc, "Thermal design studies of highpower heteroj unction bipolar transistors", IEEE Trans. Electron Devices, vol. 36, no. 5, pp. 854863, May 1989. [Get78] I.E. Getreu, Modeling the Bipolar Transistor. Elsevier Scientific, New York, 1978. [Goo95] K. E. Goodson, M. I. Flik, L. T. Su and D. A. Antoniadis, "Prediction and Measurement of Temperature Fields in SiliconOnInsulator Electronic Circuits," ASME Journal of Heat Transfer, vol. 117, pp. 574581, August 1995. [Gra84] P. R. Gray and R. G. Meyer, Analysis and Design of Analog Integrated Circuits . Third Edition, John Wiley & Sons, Inc., New York, 1993. [Gum70] H. K. Gummel and H. C. Poon, "An integral charge control model of bipolar transistors", The Bell System Technical Journal, pp. 827851, MayJune 1970.
PAGE 156
150 [Joy70] R. C. Joy and E. S. Schlig, "Thermal properties of very fast transistors," IEEE Trans. Electron Devices, vol. ED17, no. 8, pp. 586594, Aug. 1970. [Lat81] M. Latif and P. R. Bryant, "Multiple equilibrium points and their significance in the second breakdown of bipolar transistors", IEEE J. SolidState Circuits, vol. SC16, pp. 815, Feb. 1981. [Lee89] C. C. Lee, A. L. Palisoc, and Y. J. Min, "Thermal analysis of integrated circuit devices and packages,"IEEE Trans, on Components, Hybrids, and Manufacturing Technology, vol. 12, pp. 701709, Dec. 1989. [Lee92] S.G. Lee, Predictive modeling of highcurrent output resistance and thermal effects in bipolar junction transistors . Ph.D dissertation, University of Florida, Gainesville FL, Aug. 1992. [Lee95] T. Y. Lee and R. M. Fox, "Extraction of thermal resistance for fullydepleted SOI MOSFET's," in Proceedings IEEE International SOI Conference, pp. 7879, October 1995. [Lia94] M. Liang and M. E. Law, "Influence of lattice selfheating and hotcarrier transport on device performance," IEEE Trans. Electron Devices, vol. 41, no. 12, pp. 23912398, Dec. 1994. [May67] P. D. Maycock, "Thermal conductivity of silicon, germanium, IIIV compounds, and IIIV alloys," SolidState Electronics, vol.10, pp. 161168, 1967. [McA92] C. C. McAndrew, "A complete and consistent electrical/ thermal HBT model," Proc. IEEE Bipolar Circuits and Technology, pp. 200203, Oct. 1992. [Mei77] G. Meijer, "The current dependency of the output conductance of voltagedriven bipolar transistors," IEEE J. SolidState Circuits, pp. 428429, Aug. 1977. [Mun91] P. C. Munro and F.Q. Ye, "Simulating the current mirror with a selfheating BJT model," IEEE J. SolidState Circuits, pp. 13211324, Sept. 1991. [Miil64] O. Muller, "Internal thermal feedback in fourpoles especially in transistors," Proc. IEEE, vol. 52, pp. 924930, Aug. 1964.
PAGE 157
151 [Miil70] O. Muller and J. Pest, "Thermal feedback in power semiconductor devices," IEEE Trans. Electron Devices, vol. ED17, pp. 770782, Sept., 1970. [Nin84] T. Ning and D. Tang, "Method of determining the emitter and base resistance of bipolar transistors", IEEE Trans. Electron Devices., vol. ED31, pp. 409412, 1984. [Oet76] F. F. Oettinger, D. L. Blackburn, and S. Rubin, "Thermal characterization of power transistors," IEEE Trans. Electron Devices, vol. ED23, pp. 831838, Aug. 1976. [Rei92] M. Reisch, "Selfheating in BJT circuit parameter extraction," SolidState Electronics, vol. 35, no. 5, pp. 677679, May 1992. [Sof94] J. W. Sofia, "Analysis of thermal transient data with synthesized dynamic models for semiconductor devices", Proc. Tenth IEEE SemiTherm, pp. 7885, Aug. 1994. [Suh95] D. Suh, Modeling of NonFully Depleted SilicononInsulator MOSFETs, and Applications to HighPerformance/Low Power ULSI Design . Ph.D. Dissertation, University of Florida, Gainesville, 1995. [Sze81] S. M. Sze, Physics of Semiconductor Devices . Second Edition, Wiley, New York, 1981. [Sze88] V. Szekely and T. Van Bien, "Fine structure of heat flow path in semiconductor devices: a measurement and identification method", SolidState Electronics, vol. 31, no. 9, pp. 13631368, Sept. 1988. [Vog89] R. S. Vogelsong and C. Brzezinski, "Extending SPICE for electrothermal simulation, Proc. IEEE Custom Integrated Circuits Conf., pp. 21.4.121.4.4, May 1989. [Yeh95] P. C. Yeh and J. G. Fossum, "Physical subthreshold MOSFET modeling applied to viable design of deepsubmicron fully depleted SOI lowvoltage CMOS technology," IEEE Trans. Electron Devices, vol. 42, p. 1605, Sept. 1995. [Wor96] G. O. Workman, S. Krishnan, J. G. Fossum, M. M. Pelella, "DC and transient temperature dependences in SOISPICE PD and FD MOSFET models," Proc. TECHCON 96, Aug. 1996.
PAGE 158
152 [Zwe92] D. T. Zweidinger, S.G. Lee, and R. M. Fox, "Compact modeling of BJT selfheating in SPICE," Proceedings of NUPAD IV, pp. 219223, May 1992. [Zwe93] D. T. Zweidinger, S.G. Lee, and R. M. Fox, "Compact modeling of BJT selfheating in SPICE," IEEE Trans. ComputerAided Design, pp. 13681375, Sept. 1993. [Zwe95a] D. T. Zweidinger, R. M. Fox, S.G. Lee, and T. Jung, "Extraction of thermal parameters for bipolar circuit simulation," Proc. IEEE Bipolar/BiCMOS Circuits and Technology Meeting, Oct. 1995. [Zwe95b] D. T. Zweidinger, J. S. Brodsky, and R. M. Fox, "A physical thermal resistance model for vertical BJTs on SOI," IEEE International SOI Conf., Oct. 1995. [Zwe96] D. T. Zweidinger, R. M. Fox, J. S. Brodsky, T. Jung, and S.G. Lee "Thermal impedance extraction for bipolar transistors," IEEE Transactions on Electron Devices, vol. 43, no. 2, pp. 342346, Feb. 1996.
PAGE 159
BIOGRAPHICAL SKETCH David Todd Zweidinger received the BSEE degree in 1990, specializing in computer engineering, from the University of Florida. He later received the Master of Science degree in electrical engineering in 1993 from the University of Florida for his thesis work on digital electronics and digital signal processing. He is currently completing the Ph.D degree in 1997 at the University of Florida studying thermal effects in semiconductor compact modeling. He is employed at Harris Semiconductor where he is continuing work on improving circuit simulation. 153
PAGE 160
I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a thesis for the degree of Doctor of Philosophy. Robert M. Fox, Chairman Associate Professor of Electrical and Computer Engineering I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a thesis for the degree of Doctor of Philosophy. William R. Eisenstadt Associate Professor of Electrical and Computer Engineering I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a thesis for the degree of Doctor of Philosophy. . 41 L Mark E. Law Professor of Electrical and Computer Engineering I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a thesis for the degree of Doctor of Philosophy. J$T Peterson Associate Professor of Mechanical Engineering
PAGE 161
This dissertation was submitted to the Graduate Faculty of the College of Engineering and to the Graduate School and was accepted as partial fulfillment of the requirements for the degree of Doctor of Philosophy. August 1997 Winfred M. Phillips Dean, College of Engineering Karen A. Holbrook Dean, Graduate School

