|UFDC Home||myUFDC Home | Help|
This item has the following downloads:
1 CMOS MILLIMETER-WAVE RECEIV ER FRONT-END CIRCUITS AND THEIR APPLICATIONS By NING ZHANG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2009
2 2009 Ning Zhang
3 To my parents, brother, wife and my beloved baby boy, Raphael Zirui
4 ACKNOWLEDGMENTS I would like to f irst thank my advisor, Professor Kenneth K. O, for his constant encouragement and patient guidance for me through the whole chal lenging research process. I would also like to thank Dr. William R. Eisenstadt, Dr. John G. Harris, and Dr. Gloria J. Wiens for their advices and time commitment in serving on my committee. Much appreciation goes to National Science Foundation (NSF), Semiconductor Research Consortium (SRC) and Texas Instrument Incorpor ation (TI Inc.) for providing funds and access to advanced CMOS technology for our millimeter-w ave research projects. My special thanks go to Dr. Chih-Ming Hung of TI Inc. for his long-term support of chip design and fabrication, Dr. Brian Floyd of IBM T.J. Watson Research Center, a nd Dr Ivan To of Freescale Inc. for assisting our millimeter wave circuit measurements and answering mm-wave design technical questions. I have been very fortunate to work with Hsin-Ta Wu and Shashank Nallani in the mmwave projects, whose teaching and discussions have speeded up my research. I would also like to thank all my other former and current colleagues at University of Florida for their time and extreme patience to answer my questions and share their design in sights and hands-on experience generously with me. Some of their names are listed here: Chuying Mao, Yu Su, Changhua Cao, Yanping Ding, Swaminathan Sankaran, Kwangchun Jung, Dongha Shim, Haifeng Xu, Eunyoung Seok, Chikuang Yu, Zhe Wang, Xiaoling Guo, Choongyul Cha, Seon-Ho Hwang, Myoung Hwang, Wuttichai Levdensitboon, Kyujin Oh, Ruonan Han, Tie Sun, Teyu Kao, and Chieh-Lin Wu. I am deeply grateful for my parents, my brot her and my wife for their endless love, caring and encouragement, and the endless happiness my beloved baby boy Raphael Zirui brings me. This work is dedicated to them all.
5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4LIST OF TABLES................................................................................................................. ..........7LIST OF FIGURES.........................................................................................................................9ABSTRACT...................................................................................................................................14 CHAPTER 1 INTRODUCTION..................................................................................................................161.1Era of Connectivity and Scaling................................................................................... 161.2Introduction to Millimeter-Wave Technologies............................................................ 201.3Introduction to Device Technologies Supporting Millimeter-Wave ICs...................... 231.4Organization of the Dissertation................................................................................... 272 INTRODUCTION TO DIGITAL CMOS TECHNOLOGY.................................................. 292.1Nano-Scale Digital CMOS............................................................................................ 292.2Introduction to the 65-nm Digital Bulk CMOS Process...............................................362.2.1Transistor Model and Layout Optimization...................................................... 362.2.2Passive Devices in TI 65-nm Digital Bulk CMOS Process.............................. 402.2.2.1Metal-oxide-metal capacitor............................................................... 4184.108.40.206Metal-oxide-semiconductor capacitor................................................ 4220.127.116.11Spiral inductor.................................................................................... 418.104.22.168Line inductor.......................................................................................422.214.171.124Transformer........................................................................................5126.96.36.199Signal pad........................................................................................... 552.3Summary.......................................................................................................................563 CRITICAL CICRUIT BLOCKS OF MILLI METER-W AVE RECEIVER FRONT-END... 583.1Introduction................................................................................................................... 583.2Tuned Amplifier............................................................................................................ 593.2.1Design of 80-GHz Tuned Amplifier................................................................. 593.2.2Non-Ideal On-Chip Ground and Supply Network............................................ 693.2.3Substrate Coupling............................................................................................ 763.2.4Measurements of 80-GHz Tuned Amplifier..................................................... 773.3Voltage-Controlled Oscillator.......................................................................................843.3.1Design of 94-GHz Voltage-Controlled Oscillator............................................853.3.2Measurements of 94-GHz Voltage-Controlled Oscillator................................ 883.4Down-Conversion Mixer..............................................................................................913.4.1Design of 77-GHz Down-Conversion Mixer....................................................92
6 3.4.2Measurements of 77-GHz Down-Conversion Mixer........................................ 943.5Summary.....................................................................................................................1014 APPLICATIONS FOR CMOS MILLIM ETER-W AVE RECEIVER FRONT-END CIRCUITS............................................................................................................................1024.1Introduction................................................................................................................. 1024.2Introduction to 77-GHz Automobile Radar................................................................ 1024.3Overview of the 77-GHz Radar System.....................................................................1054.4Design and Measurement of an 80-GHz Down-Converter......................................... 1084.5Radio Architecture for the 77-GHz Pulsed Radar System.......................................... 1184.6Design of Receiver Chain for the 77GHz Pulsed Radar Transceiver........................ 1214.7Design of Frequency Generation System for the 77-GHz Pulsed Radar Transceiver..................................................................................................................1264.8Measurements of the Radar Receiver Front-End Integrated with the Frequency Generation System...................................................................................................... 1334.9Summary.....................................................................................................................1475 SUMMARY AND FUTURE WORK.................................................................................. 1495.1Summary.....................................................................................................................1495.2Future Work................................................................................................................151APPENDIX: SYSTEM LEVEL STUDY FOR WIRELESS INTERCHIP INTERCONNECT IN MM-WA VE BAND........................................................................................................153A.1Interconnect Bottleneck..............................................................................................153A.2Wireless Interconnect.................................................................................................. 155A.3Wireless Interconnect at 60-80 GHz........................................................................... 157A.4System Architecture and Specification....................................................................... 158A.5Low Loss Propagation Path and Efficient Antenna in a Package............................... 162A.6Initial System Specification a nd Simple Link Budget Analysis................................. 165A.7Radio Architecture......................................................................................................167A.8Duplex Scheme and System Architecture Reconsideration........................................ 171A.9Alternate Architecture with Freq uency Division Duplex Scheme.............................. 176A.10Summary.....................................................................................................................179LIST OF REFERENCES.............................................................................................................181BIOGRAPHICAL SKETCH.......................................................................................................191
7 LIST OF TABLES Table page 1-1 Some characteristics of emerging wireless technol ogies in the United States .................. 19 1-2 Some mm-wave applications below 100 GHz................................................................... 23 2-1 Capacitances of a series of m etal-oxide-metal cap............................................................ 43 3-1 Performance comparisons of CM OS mm-wave low noise amplifiers ............................... 84 3-2 Performance comparisons of CMOS fundam ental mode voltage-controlled oscillators around 94-GHz.................................................................................................91 3-3 Performance comparisons of stand-al one CMOS mm-wave down-conversion mi xer.... 100 4-1 System specification for the pr oposed 77-GHz pulsed radar system.............................. 106 4-2 Pulsed radar system range calcu lation for long distance detection.................................. 107 4-3 Pulsed radar system range calculation for short distance detection................................. 107 4-4 Component values of the tuned am plifier in the 80-GHz down-converter...................... 110 4-5 Component values of the pseudo differe ntial mi xer in the 80-GHz down-converter...... 110 4-6 Comparisons of W-band CMOS receivers...................................................................... 117 4-7 Estimation of 27.5-GHz local oscillati on leakage at the input of 27.5-GHz downconversion mi xer.............................................................................................................. 125 4-8 Performances of CMOS phase-locked loop above 50 GHz with fundame ntal VCO......138 4-9 Summary of power consumption of the frequency generation system............................ 139 4-10 Frequency generation performa nce summary.................................................................. 140 4-11 Summary of power consumption of the 83-GHz receiver front-end............................... 145 A-1 Some possible solutions for the wire interconnection bottle-neck.................................. 154 A-2 Comparison of several lo w order mo dulation schemes................................................... 161 A-3 System specification for the frequency division multiple access wireless interconnect system around 60 GHz .....................................................................................................166
8 A-4 Link budget analysis for one wireless I/ O path for a FDMA wireless interconnect system operating around 60 GHz.....................................................................................166 A-5 Wireless inter-chip interconnect system specification reflecting the adaptation of FDD scheme..................................................................................................................... 179
9 LIST OF FIGURES Figure page 1-1 Anywhere networks co nnecting our daily life............................................................... 16 1-2 Some representative wire less applications below 6 GHz.................................................. 18 1-3 Unlicensed bands assigned around 60 GHz in several countries/regions.......................... 20 1-4 Atmospheric absorption vs. frequency..............................................................................21 1-5 Added absorption due to precipitation vs. frequency........................................................ 21 1-6 Circuits operating frequency and output power range in mode rn day semiconductor technologies for high fr equency applications.................................................................... 26 2-1 Improved fT and NFmin of MOSFETs with CMOS scaling................................................30 2-2 Multiple metal layer stacking structure below protective pass ivation layer in the back-end of modern nm-range digital bulk CMOS process.............................................. 33 2-3 NMOS FET model............................................................................................................. 37 2-4 Optimized NMOS FET layout for higher fmax and ro in a TI 65-nm digital bulk CMOS process...................................................................................................................40 2-5 Metal-oxide-metal capacitor.............................................................................................. 42 2-6 Varactor layouts........................................................................................................... ......44 2-7 Spiral inductor............................................................................................................ ........47 2-8 High frequency structure simulator simulate d results for a nomi nal spiral inductor of ~250 pH.............................................................................................................................49 2-9 Various line inductor structures......................................................................................... 50 2-10 HFSS simulated results for vari ous line inductor structures (40 m long and 1.6 m wide metal6 line) in Figure 2-9.......................................................................................... 52 2-11 Baluns.................................................................................................................... ............54 2-12 Simulated |S11| and |S12| of a loop balun tuned to 50 GHz and a line balun tuned to 80 GHz from HFSS two-port simulations.............................................................................. 55 2-13 Signal pad structure and its model..................................................................................... 56 3-1 Schematic of the 80-GHz 6-stage tuned amplifier............................................................. 60
10 3-2 Comparison between common source and cascode 1-stage amplif ier around 80 GHz..... 61 3-3 CS and cascode amplifier performan ce sensitivity with device body/substrate resistance..................................................................................................................... .......62 3-4 Stagger tuning of six amplification stages......................................................................... 65 3-5 One-stage CS amplifier schematic used to tune frequency and find Sopt to obtain a good tradeoff among power gain, no ise figure, and stability............................................ 66 3-6 ZSopt, availabe power gain circles, noise fi gure circles and source stability circles around 80 GHz for low noise matching network design................................................... 67 3-7 Schematic used to design the inter-stage matching networks............................................ 68 3-8 Parasitics in the current paths including on-chip/offchip circuits.................................... 70 3-9 Schematic of a simplified on-chip gr ound parasitics network (in dashed box) connecting the sensitive nodes in the tuned amplifier. ...................................................... 73 3-10 Esitmation of ground parasitics in fluence on amplifier perform ance................................ 74 3-11 Substrate network......................................................................................................... ......77 3-12 Die photo of the 80-GHz 6-stage tuned amplifier............................................................. 78 3-13 S-parameter measurement set-up for the 80-GHz tuned amplifier .................................... 78 3-14 Measured S-parameters with 1.2 V supply and drained current of 27 mA ........................79 3-15 Noise figure measurement set-up for the 80-GHz tuned amplifier.................................... 80 3-16 Measured power gain and noise figure with 1.2 V supply, bias current 27 mA and those with 1.5 V supply, bias current 33 mA..................................................................... 81 3-17 Linearity measurement set-ups.......................................................................................... 82 3-18 W-band signal sources..................................................................................................... ..83 3-19 Measured IIP3 and IP1dB of the amplifier under 1.2 V supply and 27 mA drained current................................................................................................................................83 3-20 Schematic of the 94 GHz voltage-controlle d osci llator with all transistors sizes in m...86 3-21 Part of the cross-coupled transi stor pair layout with spacings in m. ...............................86 3-22 Inductor and vara ctor layouts.............................................................................................87 3-23 Die photo of the 94-GHz VCO..........................................................................................88
11 3-24 Measurement set-up for the 94-GHz VCO........................................................................ 89 3-25 Measured VCO output spectrum before de-embedding.................................................... 89 3-26 Measured phase noise with 1.5 V and 6 mA VCO core supply with carrier frequency of 94.92 GHz................................................................................................................... ...89 3-27 Voltage-controlled oscillator tuning curves with 1.5 V supply and bias currents of 2, 4, 6 mA for the VCO core.................................................................................................. 90 3-28 Schematic of the 77-GHz mixer...................................................................................... 942 3-29 Die photo of the 77-GHz mixer......................................................................................... 94 3-30 Measured ports matching of the mixer with 1.2 V supply and 5 mA bias current............ 95 3-31 Measurement set-up for the 77-GHz mixer conversion gain............................................. 95 3-32 Measured conversion gain of the mixe r under varying bias currents and LO power levels......................................................................................................................... .........96 3-33 Measured average mixer conversion gain (CG), SSB NF and NF after de-embedding RF input balun loss (NF2)..................................................................................................97 3-34 Measurement set-up for the 77-GHz mixer NF................................................................. 97 3-35 Measurement set-up for the 77-GHz mixer IIP3 and IP1dB................................................98 3-36 Measured IIP3 with two input tones of 76.3 and 76.31 GHz and IP1dB with an input at 76.3 GHz under 1.2 V supply a nd 5 mA drain current......................................................98 3-37 Measurement set-up for the 77-GHz mixer LO-RF leakages............................................ 99 3-38 Measured leakages among mixer ports at VDD=1.2 V and 5 mA drain current................. 99 4-1 Vehicle with multiple radar modules for va rious functions to ma ke driving more safe and comfortable...............................................................................................................103 4-2 Multi-beam antenna and its radiation pa ttern for the proposed 77-GHz CMOS radar system......................................................................................................................... .....105 4-3 Schematic of the tuned amplif ier in the 80-GHz down-converter................................... 109 4-4 Schematic of the pseudo differentia l mi xer in the 80-GHz down-converter................... 109 4-5 Die photo of the 80-GHz down-converter....................................................................... 112 4-6 Measured and simulated impedan ce ma tching of the down-converter............................ 114
12 4-7 Measured conversion gain and NF of th e down-converter for fixed local oscillation frequency of 60 GHz........................................................................................................ 115 4-8 Measured conversion gain of the down-c onverter at varying pow er levels of the 60GHz LO deli vered to the LO port.................................................................................... 115 4-9 Measured IIP3 and IP1dB of the down-converter at 81 GHz............................................ 116 4-10 Measured LO-RF and LO-IF l eakages of the down-converter........................................ 116 4-11 Simplified block diagram of the CMOS he terodyne radar transcei ver showing the LO frequencies generation scheme and separate on-chip dipole antennas for TX and RX... 118 4-12 Receiver block diagram showing the main 25.5-GHz LO leakage paths........................ 119 4-13 Three-dimension modeling of the transition from the 76.5-GHz on-chip dipole feed to the waveguide section of a high gain antenna............................................................. 120 4-14 Receiver mm-wave front-end schematic (R F stages) of the 77-GHz radar system......... 122 4-15 Receiver IF and BB stages sche ma tic of the 77-GHz radar system................................ 123 4-16 Block diagram of the 51-GHz PLL frequency synthesizer (in dash box) in the proposed 77-GHz radar transceiver................................................................................. 126 4-17 Latch-biased static frequency divider used for the 1st and 2nd stage of divider............... 128 4-18 Simulated input sensitivity plot of the 1st latch-biased static frequency divider............. 131 4-19 Die photo of the 77-GHz radar frequency generation system......................................... 134 4-20 Measurement set-up for the 77-GHz radar frequency generation system....................... 135 4-21 Frequency tuning and singl e-ended output power of VCO and offset mi xer after deembedding the measurement set-up loss.......................................................................... 135 4-22 Spectrum of the phase-locked loop 55.5 GHz output...................................................... 136 4-23 Phase noise of the PLL 55.5 GHz output......................................................................... 137 4-24 Up-converted 83.3 GHz spectrum at the offset mi xer output corresponding to 55.5 GHz output of the PLL....................................................................................................137 4-25 Phase noise at 1 MHz o ffset and reference spur.............................................................. 138 4-26 Die photos of the .............................................................................................................141 4-27 Measurement set-up for the receiver front-end................................................................ 142
13 4-28 Measured and simulated input matching |S11| of the receiver front-end.......................... 143 4-29 Measured and simulated conversion gain and SS B NF of the receiver front-end........... 144 4-30 Measured IIP3 and IP1dB of the receiver front-end........................................................... 145 A-1 Wireless interconnects help reduce multi-chip p ackage area and I/O wiring complexity..................................................................................................................... ...156 A-2 Architecture of an frequency division multiple access wireless inter-chip interconnection system.....................................................................................................160 A-3 Multiple chips/components in an multi-chip module...................................................... 162 A-4 Enhanced wave propagation environment for an MCM.................................................. 163 A-5 HFSS simulated input return loss of a bond wire antenna around 60 GHz. .................... 164 A-6 A series of bond wire antenna test stru ctures consisting of silicon chips, gold bond wires and bond pads on FR-4 board................................................................................ 164 A-7 Measured antenna pair gains of bond wire antennas with and without me tal cover....... 165 A-8 Radio circuit architectures............................................................................................... 169 A-9 A bilateral wireless FDMA interconnect in time division duplex schem e is realized with three RF transceivers and bond wire antennas on each chip................................... 171 A-10 Adjacent bond wire antenna coupling cause s serious adjacent ch annel interference problem and degrades received signal SNR. ................................................................... 172 A-11 Adjacent bond wire antenna coupling causes serious reciprocal mixing problems and degrades received signal SNR......................................................................................... 174 A-12 An alternative wireless FDMA interconnect architecture increases hardware sharing and m itigates undesirable bond wire antenna coupling................................................... 175 A-13 Wireless FDMA interconnect system usi ng an FDD scheme utilizes tw o frequency bands for bidirectional communication, removes the T/R switch and reduces undesirable bond wire antenna coupling.......................................................................... 177 A-14 Multiple buffers, notch filters and offset mixers can be added to the 60-GHz band frequency synthesizer in Figure A-8 (B) to generate the 80-GHz band LO signals needed by the FDD scheme .............................................................................................178
14 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy CMOS MILLIMETER-WAVE RECEIVER FRONT-END CIRCUITS AND THEIR APPLICATIONS By Ning Zhang August 2009 Chair: Kenneth K. O Major: Electrical and Computer Engineering. The improvement of high frequency capability for silicon devices has made implementation of millimeter-wave (mm-wave) sili con integrated circuits operating at 60 GHz, 77 GHz and even higher feasible. This had led to the proposal of a low-cost 77-GHz CMOS transceiver for automobile radar application and a 60-GHz wireless inter-chip interconnect system. This Ph.D work demonstrated an 80-GHz mm-wave receiver chain integrated with a frequency synthesizer fabricated using low leak age transistors of a low cost 65-nm bulk CMOS technology. An 80-GHz single-ended low noise amplif ier (LNA), a 77-GHz down-conversion mixer and a 94-GHz voltage-controlled oscillator (VCO) have been separately demonstrated. A 1.2-V supply, the single-ended LNA exhi bits 12-dB gain and 9-dB noi se figure (NF) consuming 32mW power. The mixer has -dB conversion ga in and 17 dB NF while consuming only 6-mW power. With a 1.5 V supply, the VCO achieve s 5.8% tuning range around 94 GHz and 87dBc/Hz phase noise at 1 MHz offset while consuming 14-mW power. These designs demonstrate that the low-cost, low leakage CMOS process can be used for the design of mmwave circuit blocks and potentially larger integrated system desp ite the challenges of using the
15 technology such as low voltage headroom, moderate metallization performance and strict metal density filling requirements. An 80-GHz down-converter including an LNA and a mixer is demonstrated. It achieves 16-dB conversion gain and 9-dB noise figure, consuming 52-mW power from a 1.2-V supply. To build a more complete radar re ceiver system, a signal generation system using a phase lock loop is demonstrated. It generates 55 to 60 GHz a nd 27.5 to 30 GHz LO signals for the receiver downconversion, and 83 to 90 GHz carriers for the transmitter. The signal generation system consumes ~170 mW power with the supplies of 1.2 and 1.4 V. This work has shown that the low-cost CMOS technology can be used to build wide-locking range, low phase noise and low power consumption mm-wave signal generation circuits. Finally, a receiver chain is integrated with the frequency gene ration system to demonstrate a radar receiver. The receiver chain starts w ith an 80 GHz LNA using an on-chip balun to convert single-ended input signal to differen tial, a mixer following the LNA down-converts the incoming signal at 80-GHz to 27-GHz, then an IF amplifier and passive mixer amplify and down-convert the 27-GHz signal to baseband. Lastly the baseband signal is high-pass filtered and output to an external load through a s ource follower buffer. With 1.4/1.2 V power supplies and LOs generated on-chip, the receiver circuit has about 7-dB conversion gain, and NF is as low as 13.8 dB. The IP1dB and IIP3 are and dBm, respectively. The goal of this research is fulfilled. We successfully demonstrate the feasib ility of realizing a low-cost single chip mmwave transceiver for automobile radar applications at 77 GHz and beyond.
16 CHAPTER 1 INTRODUCTION 1.1 Era of Connectivity and Scaling Thanks to the rapid progress of electronics, computer and network tec hnologies in the past 50 years or so, we are now living in an electronic inform ation age. All information is digitized and spreaded through a global network formed by numerous optical /coaxial backbone, microwave and satellite relay a nd cellular/ad-hoc access. Data c onnection, from long distance to local area network (LAN) , and lately pers onal area network (PAN)  and body area network (BAN) , is interwoven with peoples da ily life, providing conve nience and productivity enhancement. Figure 1-1 is an illust ration for this anywhere connection. Figure 1-1. Anywhere networ ks connecting our daily life Observing the development of integrated ci rcuits (IC) industry in the background of the electronic information age, we see that user demands, market volume, technology investment
17 and development closely couple with each othe r, forming a positive feedback loop: User demands and existing needs stimulate capitalists and engineers to look for technical solutions, huge return/loss prepare people for a cycle of problems and solutions. On the other hand, technology breakthroughs lead to ideas for new applications and commercialization. This is the driving force for the advancement of the whole elect ronics industry. Particularly, it has fueled the Moores law, which has become the gauge for elect ronics to measure the pace of very large scale integration (VLSI) integrated circui ts technology evolution . Users always look for cheaper and faster data services and more functions in a smaller device. This thirst drives for faster com putation, larger storage and wider communication bandwidth , pushing the whole electronics industry to take every possible means to continuously lower design/manufacture cost of pr oduct while squeeze in more functions. Thanks to these efforts in material/p rocess and circuit arch itecture/topology innovations, and relentless scaling-down of semiconductor device technology feature sizes, now we posses relatively lowcost microprocessors with clock rate of GHz , Giga-bytes memory  and Giga-bps wireline communication ICs . However, the bottleneck of communications still exists particularly in the area of wireless communication and multi-GHz rate interconnects such as processor-toprocessor or processor-to-memory. This bottleneck also explains the proliferation of research activities in the area of wireless/wireline ICs in the past 20 years or so. The commercially available cellular phone system-on-a-chip (S OC) with multi-mode, multi-band, self-calibrated, digitally assisted RF/analog circui try exemplifies one of the latest efforts and directions of the wireless communication IC industry, which was sti ll just a concept 10 year s ago , . The ever-increasing sampling speed of analog-to-dig ital conversion and the concept of transforming the voltage domain signal processing to time dom ain processing ,  could ultimately lead
18 to the true software-defined-radio (SDR)  th at could change the f unction of an extremely versatile communication hardware platform in a blink of software program loading. Due to the mobility, flexibility and relatively easy, cheap deployment it brings, wireless communication technology has become an indispen sable part of the global data network and a significant portion of the global el ectronics industry , . Th is has let to a need for highspeed wireless technology. However, to attain a wireless link w ith data rates of ~100 Mbps-1 Gbps and beyond, there are two fundamental obstacles that must be overcome. Figure 1-2. Some representative wireless applications below 6 GHz The first one is the available communicat ion bandwidth for wireless applications. Nowadays, it is a global norm for governmental organizations to assign or license limited frequency spectrum to different services/applicat ions. Use of the spectrum in a manageable fashion avoids chaos and mutual interferences. However, multiple factors such as the availability of low-cost technology, lower propagation path loss, industrial expertise and investment accumulation make the spectrum below 6 GHz already packed with all sorts of consumer
19 applications with a variety of communication distance and inform ation rate , . Wireless local-area-network (WLAN) and cellular phone are prob ably the two examples that many people have experienced. Figure 1-2 show s a diagram for some representative applications below 6 GHz. The maximum data rates of all these applic ations may be ultimately limited by the narrow bandwidths they are assigned and th e strong interferences they mutu ally cause. Several technical solutions have been proposed to circumvent these issues : ultra-wide-band (UWB) technology, cognitive radio and mm-wave techno logy (frequencies from 30 GHz to 300 GHz). Table 1-1 summarizes typical ch aracteristics of these technol ogies in USA. Among them, UWB and mm-wave both can satisfy the need for modera te-to-high data rate applications. UWB faces the challenges of stricter transmitting power le vel control and existence of strong narrow-band interferences from other applic ations in the populated band. Thes e limit UWB to short distance (<10 m) applications. On the other hand, mm-wave systems have larger available spectrum resources worldwide that are relatively clean since fewer applications exist there presently, and the transmission power level can be much higher than UWB . Millimeter-wave transceiver modules for Gbps point-to-point communication for ~1 km range are commercially available . Figure 1-3 shows the much noted 60 GHz unlicensed band spectrum designated in several developed regions in the world . Such rarely seen global harmony attracts the interest of industry and it could well be the start of anot her great success story following the world-wide acceptance of IEEE 802.11 WLAN. Table 1-1. Some characteristic s of emerging wireless techno logies in the United States Technology UWB radio 60 GHz radio Cognitive radio Spectrum access Underlay Unlicensed Overlay Spectrum (GHz) 0-1 & 3-10 57-64 0-10 Data rate 100-500 Mbps > 1 Gbps 10-100 Mbps Range ~1-10 m ~1 m-1 km ~1 m-10 km
20 Figure 1-3. Unlicensed bands assigned ar ound 60 GHz in several countries/regions The second obstacle is the need for an affordable semiconductor technology that not only provides adequate and reliable pe rformances for ~Gbps wireless ci rcuits, but also suits large volume production and scale of economy. Fortunate ly, benefiting from ceaseless scaling down of semiconductor devices, the operating speed of them has been soaring, which will make this obstacle to be less of a problem in the next few years even for circuits operating up to 200 GHz. One hidden advantage of mm-wave technology is that it is relatively easy to form a complete mm-wave transceiver by adding a mm-wav e front-end in front of the lower frequency (below mm-wave band) RF circui ts or just replacing the RF front-end with an mm-wave frontend. As a rich body of knowledge has been accumu lated for lower frequency RF circuits and systems over the years, it is possible to expl ore the feasibilities of implementing the proven circuits and architectures at mm-wave frequencies without having to start from the scratch. With such an idea, an architecture where 60 GHz fr ont-end and 5 GHz front-end are cascaded to form a reliable dual WLAN system has been proposed . 1.2 Introduction to Millim eter-Wave Technologies Systematic research of mm -wave dates back to 1950s, when people were first able to build coherent generators for mm-wave frequencies . Ever since then, this frequency band has long been used for material spectroscopy, atmosphere/astronomical spectroscopy, military radar and space communications etc. Bulky, high-power vacuum devices such as magnetron,
21 klystron etc, were used as mm-wave power sour ces and extremely high gain antennas were used to obtain sufficient power densities radiated to sp ace . With the develo pments of scientific and military applications, the propagation charac teristics of mm-wave were studied thoroughly (Figures 1-4 and 1-5) . 60 GHz ~18 dB/km 77 GHz ~0.5 dB/km Figure 1-4. Atmospheric abso rption vs. frequency. Figure 1-5. Added absorption due to precipitation vs. frequency. As can be seen in Figure 1-4, in the mm-wave band, the excess loss (in addition to the free space path loss predicted by Friis formula) shows peaks and valleys in log scale, which is particularly large at some frequencies due to the absorption/resonance of certain components of the atmosphere. E.g., the oxygen absorption at 60 GHz results an excess loss of ~18 dB/km in
22 comparison to the rather low excess loss of ~0.5 dB/km at 77GHz (both measured at sea level). This loss is exacerbated further by weather condi tion as shown in Figure 1-5, where an extra attenuation factor depending on the level of prec ipitation is added to the attenuation levels of Figure 1-4. This additional loss sp ecially influences system de sign considerations for outdoor mm-wave applications such as vehicu lar radars and point-to-point links. At the same time, due to the small wavelengt hs (1 mm~10 mm) of mm-wave signals, their propagation behavior is cl oser to line-of-sight (LOS) and there is less diffraction compared to the signals at lower RF frequencies. The mm-waves also experience a larger loss from reflections and scattering. The above characteristics of mm-waves in co mbination with the relatively clean and large bandwidths could potentially be tu rned into advantages in numerous applications. Millimeterwave radar and imaging both utilize the mm-w ave properties of shor t wavelengths, high directivity and reflections/sca ttering behaviors to achieve hi gh-resolution detection. Another example is that the high loss of 60-GHz band actua lly makes this band suitable for short-distance (e.g., in-door) high data rate commu nications. With proper networ k planning, efficient frequency reuse and high user density (system accommodated simultaneous user number in a certain area) could be achieved. In contrast the relatively low loss of 80-100 GHz band makes it more desirable for point-to-point data links that could extend the co mmunication distance to ~km. At the same time, the propagation channel multi-path dispersion problem, which is common in wireless communication, could be mitigated by the quasi-LOS propagation and relatively large loss of reflection/scattering of mm-waves. On the other hand, the higher pa th loss and excess loss of mm-waves may lead to larger power consumption. The quasi-LOS propagation property
23 reduces the spatial coverage. Such disadvantages could be mitigated by system and circuit level optimization. 1.3 Introduction to Device Technol ogies Supporting Millimeter-Wave ICs In the last 4 years, we have witnessed a soaring interests in mm-wave technologies in both industrial and academic institutions, along with the emergence of mm-wave Gbps link and WPAN, wireless high-definition multimedia interface (W-HDMI), automotive radar and mmwave imaging etc. The reviving interests in mm -waves are stimulated by the large available spectrum, potentially huge market and aggressi ve technology advancement. Some of the mmwave applications below 100 GHz with large co mmercial potential are listed in Table 1-2. Table 1-2. Some mm-wave applications below 100 GHz Applications Frequencies Broadband wireless data link 60 GHz band (unlicensed) 71-76 & 8186 GHz band (licensed) Satellite communication down-link 26-40 GHz Vehicle-vehicle co mmunication 63-64 GHz Automobile radar 76-77 GHz Imaging 94 GHz Before discussing the main semiconductor tec hnology available for mm-wave applications, it should be pointed out that for RF and mm-wave circuit systems, the package and testing costs are a very large chunk of the overa ll expenses and may even be th e dominant. In this context, integration level of an IC technology could be th e key to cut the cost: an IC which can include most circuit blocks and components from RF to analog, providing self-calibrating, self-test and signal processing as a result of th e integration of VLSI digital circuitry will surely ease the system package and testing complexity, lowe r component counts, a nd package area, hence reduce the overall system cost.
24 Looking back, as early as 30 years ago, solid-s tate circuits and systems operating in the mm-wave range were demonstrated . The so lid-states mm-wave circuits were first built using III-V semiconductor, where the two most wi dely used ones are Gallium Arsenide (GaAs) and Indium Phosphide (InP). Due to intrinsic higher mobility of carriers and lower-loss substrate, fT (unity current gain frequency) and fmax (unity power gain frequency) of active devices in III-V technologies have been leading those of silic on technologies. Today, the state-of-art InP HBT technology already deliver fT up to 400 GHz and fmax possibly close to 1000 GHz . In addition, the usually thicker back-end metal and la rger distance to the low-loss substrate of III-V result high-quality passives. III-V technologies have suffi cient performance for mm-wave frontend circuits and -[28 ] are some examples of this. Neverthe less, the cost of III-V technologies has been much higher than silicon-based te chnology (Si-technology) due to the higher wafer cost, smaller wafer size and lower yield. Moreover, III-V IC technologies usually have a fewer back-end metal layers compared to Si-technol ogy, which reduces the achievable integration level. And it is very challenging to make II I-V technology and Si-tec hnology co-exist on the same die. Hence, in order to build a complete radio system utilizing the computing power offered by modern Si VLSI technology to process the do wn-converted signals from III-V front-ends, we have to package multiple dies together in the same package, which adds cost and degrades the reliability of the system. Because of this, it is difficult for III-V technologies to become the mainstream technology in the mass consumer ap plication where Si-based technologies are already able to handle. This k eeps the III-V to be low volume, less available, low yield and high cost technologies. With the fT and fmax of Si-technologies alr eady at 100-300 GHz range , and expected to be highe r in the near future, replacing III -V with Si to build low-cost,
25 moderate performances mm-wave front-ends fo r consumer applications is an exciting opportunity. Silicon integrated circuits technology has been the most powerful engine for the semiconductor industry growth over the past 50 year s -, especially in the last 30 years, with the maturing of silicon CMOS VLSI technology. Abundant substrat e material, excellent mechanical and thermal characteristics, easily fo rmed excellent isolation (silicon dioxide), are just some of the obvious adva ntages for Si technologies. There are mainly two types of Si technologies, bipolar and CMOS. Both achieves higher speed with technology scaling and ca n integrate millions of transist ors at small incremental cost with superior matching, repeatability and hi gh yield at ever incr easing wafer size. The bipolar technology for high fr equency operations is mostly Silicon Germanium (SiGe) hetero-junction bipolar transistor (HBT) technology and it is often offered in a BiCMOS process. The latest SiGe technology has been able to o ffer high frequency performance comparable to IIIV technologies ,  at lowe r cost. The possibility of making silicon circuits operate at mmwave frequencies has also b een demonstrated -. Comparing to CMOS, due to th e inherent device structure di fferences and the enhancement of hetero-junction technology, SiGe devices has higher carrier mobility, lower flicker noise, and higher breakdown. However, SiGe technology sti ll costs more than CMOS at the same technology node because of more complex processing steps needed. Although SiGe HBTs in BiCMOS technology with a coarser lithography (and hence lower cost) can offer higher frequency performance than CMOS at a more advanced node requiring finer lithography, MOS transistors in the same BiCMOS process is lim ited by the coarser lithog raphy. This translates into limited digital performances and lower integration level (hence larger area).
26 CMOS no doubt is the dominant IC technology today: Its superb switching and low power characteristics, and relatively simple processing (such as self-aligned gate formation) have paved its way to the unparallel success. Th e success of CMOS in Gbps wire line serial-link circuits and low GHz RF circuits in the past 10 years has been impressive. L eaping into the deep sub-micronscale and nano-scale era, CMOS, which once was considered only suitable for low frequency circuitry, is already showing the promise for being able to handle at mm-wave frequencies , -. Digital bulk CMOS technology, as the lowest co st and highest integration solutions among all, understandably, was chosen for RF app lications in ITRS 2005 . Although the high development cost of nano-scale digital CMOS is a concern for RF/mm-wave applications, CMOS will eventually prevail in large volume c onsumer markets, especially where a high level of integration is desired to lower system cost and enhance functionality. With this vision, exploring the mm-wave capabilities of a low-cost digital bulk CMOS technology is a necessary first step. Figure 1-6. Circuits operating frequency and output power range in modern day semiconductor technologies for high frequency applications
27 Assuming that circuits in a given semiconduc tor technologies can ope rate up to one third of the peak fmax of active devices, Figure 1-6 shows the approximate ranges for operating frequency and output power in different technologies discussed above. 1.4 Organization of the Dissertation This thesis research focuses on the d esign and characterization of mm-wave circuits and systems in a low-cost 65-nm CMOS technology. Th e goal is to demonstrate the feasibility of implementing mm-wave circuits using low-leak age transistors in the low-cost bulk CMOS. In Chapter 2, after reviewing some general tr ends of nano-scale CMOS which are relevant to mm-wave circuit design, the low-cost 65-nm CMOS technology being used for this research is introduced. The active device model and some key passive devices are presented, with the discussions on component de sign and modeling issues. Chapter 3 then follows with the presentations of several basic buildi ng blocks of mm-wave receiver front-end implemented using low-leak age transistors of the 65-nm CMOS technology. The designs and characterizations of an 80-GHz tuned amplifier, a 94-GHz voltage controlled oscillator, and a 77-GHz down-conversion mixer are described in sequence. In Chapter 4, following the introduction of the 77-GHz automobile radar application background, the implementation of an 80-GHz down-converter is discussed and its performance is presented. Next, a pulsed ra dar transceiver system in CMOS is proposed, the design of a 77GHz receiver front-end including re ceiver chain and frequency generation system is reviewed. Finally, the measured data of the frequency generation system and the receiver chain are presented. Chapter 5 summarizes the thesis and discu sses possible future wo rks to improve the performance of the 77-GHz receiver front-end.
28 Following Chapter 5, the system-level study of a wireless inter-chip interconnect system operating at mm-wave band is reviewed in th e appendix, as another example of possible application of CMOS mm-wave circ uits. A wireless interconnect sy stem architecture is proposed based on the application scenarios. The key system specifications and circuit blocks specifications are discussed.
29 CHAPTER 2 INTRODUCTION TO DIGITAL CMOS TECHNOLOGY 2.1 Nano-Scale Digital CMOS Silicon scaling has push ed the digital CMOS technology into the nano-meter era. The scaling no longer singly relies on just gate length (Lg) reduction. New materials, processes and device structures are incorporated to meet th e necessary performance requirements. Device mobility enhancement with cha nnel stressing and use of non <100> orientation substrates has gained wide acceptance . Novel device structures such as multiple gate transistors are also envisioned to provide stronger channe l control and larger transconductance (gm). RF/mm-wave applications also benefit from such logic-dr iven scaling and innovati ons. MOS field-effect transistors (FETs) expect improved fT and NFmin with every new generation. Figure 2-1, an excerpt from  clearly shows this improving trend for every generation down. The data are for radio-frequency (RF) CMOS and silicon-on-in sulator (SOI) CMOS te chnologies. Although not shown in the plot, digital bulk CMOS follows a similar trend. However, digital CMOS evolution down to the sub-100nm range brings a new set of challenges for designing especia lly at high frequencies can count eract the scaling benefits. We need a keen understanding of so me phenomena relevant to mm-wa ve circuit design and look for approaches to overcome the challenges: First, high gate and sub-threshold leakage of nano-scale MOSFET plague the CMOS VLSI circuit with high static power consumption and heat dissipation. In lowcost digital CMOS, a straightforward solution for these leakages is to increase threshold voltage of the devices and reduce the supply voltage. The higher threshold will lower the current driving capability and hence the high frequency performance. The lowe r supply voltage helps to implement constant field scaling , reducing risk of breaking thinner gate oxide. However, the voltage headroom,
30 the linearity of circuits, and the maximum out put power capability of a single device are degraded. To mitigate the impact of high thre shold voltage and the low voltage supply, the RF/mm-wave circuit topology could opt for simp le, less stacking and more folded styles. To generate the necessary output power, power co mbining with a planar power combiner, or a distributive transformer, or spatial comb ining with antenna array may be used. Figure 2-1. Improved fT and minimum noise figure of MOSFETs with CMOS scaling . Second, the gate oxide of nano-scale MOSFET (<3 nm) has been thinning along with the shortening of the channels so that the gate can retain control over the channel (switching on/off). However, this leads to lower breakdown and highe r possibility of electro n tunneling. To alleviate these, a high K gate dielectric material is us ed to slow down the pace of gate oxide thickness shrinkage while maintaining the effective cha nnel control. Meanwhile, adding strains to the channel helps raising carrier mobility but that also brings in the dependence of mobility on active
31 device layout finger width and le ngth . High verti cal field induced by gate voltage causes mobility degradation and hence reduces gm in saturation region, as indicated by Equation 2-1. )(1 )(TGS n GSeffVV V (2-1) describes mobility degradation due to the vert ical electric field , and n is the low field mobility. Equation 2-1 indicates there is an optimum gate control voltage for maximum gm in saturation. The classical square-law for device current in saturation is no longer valid, as seen in Equation 2-2. )1( )](1[2 )(2 DS TGS ox TGSoxn DsatV VVLt VVW I (2-2) Although gm keeps increasing from one generation to the next, because the channel length modulation effect also increases, l eading to lower output resistance ro and intrinsic gain gmro. Along with the complex bulk substrate resistive ne twork, the low output re sistance is a serious limiting factor for circuit gain. To mitigate this issu e, the body resistance of the transistor can be increased by reducing the number of transi stor substrate contacts in the layout. The equivalent non-quasi-static (NQS) channel re sistance has a larger effect at mm-wave frequencies compared to at low GHz frequencie s. This can degrade the input matching quality factor. Third, high horizontal field effects in the extremely short channe l (<100 nm) such as velocity saturation and drain induced barrier lowering (DIBL) lim it the current drive and damage the control of gate voltage on channel. e.g., th e threshold voltage varia tion is found to be ~50mV when drain voltage changes between 0 and 1 V for a 90-nm NMOS device . Solutions like non-uniform channel doping (e.g., halo impl antation) and lightly-doped-drain (LDD) are applied to mitigate the drain control over the channel. However, the drain depletion width
32 reduction associated with halo implantation fu rther exacerbates the br eakdown voltage of the drain junction. Fourth, there has been concern that the hi gh field in the channel could produce more carriers scattering and hence add excess thermal noise at high frequencies. However, so far CMOS scaling seems to still improve device noise figure at low frequencies partially because the shortening channel reduces the possibility of scattering events . Flicke r noise also benefits from the similar mechanism. Also, low quality passives could contribute significant noise. Hence matching active devices to the mi nimum noise figure at its input does not n ecessarily produce the lowest noise figure for the whol e circuit. Trading the optimum device noise matching for lower loss of smaller matching elements could lower overall noise figure. Successful RF/mm-wave designs rely on perfor mances of both active and passive devices. Many passive devices are formed by the back-end metal layers. The interconnects within active device layout also influe nce its high frequency characteristic s. The high frequency performance of these devices strongly depends on the quality of back-end layers. That makes the study of these back-end layers a necessity. In order to increase integr ation density of active devices, pitches and widths of interconnect lines (including pol ysilicon layer) in na no-scale digital CMOS are drastically reduced (<100 nm). To prevent mechanical cr acking of the metal layers and obtain reliable interconnection, the vertical dimensions of metal and sizes of vias/contacts are shrunk simultaneously. This forces the dielectric layer th ickness to be reduced to avoid the difficult task of building fragile tiny vias/contacts stretched long in the vertical direction. All these lead to a stacked structure that has multiple very thin metal layers and insulator layers on top of the silicon substrate as shown in Figure 2-2. The top copper metal layer and the substrate are separated by a
33 distance typically of 3-6 m and the metal layers are connected by very small vias/contacts. Naturally, such a structure results in relatively large ohmic loss from both wires and vias/contacts, and high capacitive coupling am ong metal layers and between metals and substrate, which have severe adverse effect on both active and passive devices characteristics. Figure 2-2. Multiple metal layer stacking struct ure below protective passivation layer in the back-end of modern nm-range digital bulk CMOS process (not to scale) Unfortunately, a modern digital CMOS proce ss usually employs moderately to highly doped silicon substrate, and the cl ose proximity of metals to th e substrate intr oduces undesirable substrate loss and coupling. High power applications are also deep ly affected by the thin metal layers, because the wide metal trace needed to accommodate high current implies large capacitive loading (and remember the short di stance to the substrate), which makes high frequency operation difficult. Sim ilarly, shunting metal traces on different layers will affect the high frequency operation.
34 To alleviate these issues, copper has been intr oduced to replace aluminum to obtain higher conductivity, and low K dielectrics are used as in sulator among metal laye rs to lower the metalto-metal coupling . There is also a possi bility of replacing polysilicon gates in MOSFET with metal gates in th e future nodes . Since flat surfaces of each metal layer in the digital CMOS process are needed to support nano-scale pitches, chemical mechanical polishing (CMP) is used to create such flat surfaces. To provide necessary mechanical support and achi eve uniform polishing, strict metal density requirements are imposed for each metal layers. In turn, strict metal filling rules are enforced in the layout of the chip, which not only exacerba tes the high frequency signal loss and coupling but also greatly increases the layout time. In some customized CMOS processes, more metal layers, extra thick metal layers and thicker dielectric layers are supported to allevi ate high resistance from thin metal and coupling issues. In addition, special waivers are offered to designers to avoid obeying the strict metal density requirements. However, the former adds the cost and the latter could harm the overall wafer yield. To fully exploit th e cost advantage of CMOS, it is desirable to use a pure digital process trading off some performance. On the bright side, as can be seen in Figure 2-2, the top copper layer is usually thicker for long distance interconnects in nano-scale digital CMOS. And the process usually offers a thick aluminum layer (ALCAP) above all copper layers for top level interconnection and signal pad. In addition, one or two thick copper layers fo r redistribution (RDL) purposes above chip passivation layer can be added dur ing the packaging process. If being used appropriately, these layers could help improve passive device quality factors, although having much wider minimum widths and spacing than lower copper layers make their use not straightforward. What is more,
35 nano-scale digital CMOS could support a layer named high re sistivity (HIRES), which could be laid underneath passive components and criti cal signal paths to block implantation, augment the substrate resistance and reduce substrate loss. Proper circuit design boils down to accurate modeling for bot h active and passive devices, which is truly a very challenging task at mm-w ave frequencies. The necessary measurements are prone to errors due to difficulty of calibration and de-embedding at such high frequencies. The measurement set-ups are expensiv e. Distributive effect s of devices are more pronounced at mmwave frequencies due to the shor t wavelengths on silicon chip. Many parasitic effects are more obvious at such high frequencies and the models ha ve to include more variables. Thus, simple lumped models proven to be useful at lower frequencies could be qu ite inaccurate. Relying on modern 3-D EM simulations and optimization to ols, paying attention to additional physical parasitic effects at high freque ncies, finding correct measuremen t/data processing procedure and using multiple section lumped model are just some ways to deal with the challenge. Finally but not lastly, due to the overall s caling of sizes from transistors to metal interconnects, plus the Vth variation due to DIBL and strain, the variation among devices in nano-scale CMOS process could be relatively large and the matching of devices in circuits could be degraded. This again points to simpler circuit topologies for high fr equencies. Constant current density biasing scheme that has long been used in bipolar circuits could also help to lower the influence of process variations . In summary, tremendous opportunities and tough challenges exist for mm-wave design in nano-scale digital bulk CMOS. Digital bulk CM OS provides the lowest cost and highest integration for mass-market applications. On the ot her hand, it presents a series of issues for mmwave designs such as low voltage supply, high threshold voltage, low breakdown, low output
36 resistance, thin interconnect metals and strict me tal density rules. To meet these challenges, we need to better understand the device operation and tradeoff wi th the assistance of EM simulation tools, to optimize device implementation within the constraints of desi gn rules and to look for simple and robust circuit topologi es adapted to the process. 2.2 Introduction to the 65-nm Digital Bulk CMOS Process For the mm-wave works in this proposal, Texas Instruments (TI) 65-nm digital bulk CMOS technology is used. This tec hnology features low leakage (high Vth, ~0.6 V) NMOS transistors with fmax of >110 GHz and fT of >150 GHz and 6 copper metal layers. The top copper layer is ~1.5 m thick and has a distance to the substrate of >2 m. Other copper layers have thicknesses <0.2 m. Contact resistance is ~60 / contact and vias for the lower 5 copper layers show ~4 /via. The silicon substrate has a resistivity of ~10 cm-1. The nominal supply voltage is 1.2 V and gate to drain breakdown vol tage is > 2V. The digital technology is also equipped with HIRES, ALCAP and RDL layers as mentioned in the previous section. Strict metal density rules are enforced for each copper layers. All spaces on chip except those used for circuit interconnects will be filled with floating dummy metal patterns. 2.2.1 Transistor Model and Layout Optimization For high frequency circuit design, NMOS devi ces are preferred to PMOS devices due to their higher carrier mobility. The transistor model  must include parasitics such as gate resistance (Rg), body resistance network (equivalent to Rsub), and drain/source junction capacitances (Cdb/Csb). The NMOS transistor model used fo r our designs is similar to the one shown in Figure 2-3 , . Th e core model parameters can be configured based on transistor DC I-V characteristics, and all ot her components in the model can be fitted using high frequency S-parameter measurements.
37 A straightforward way of evalua ting high frequency performance of transistor is to look at their fT, fmax and NFmin. ( fT =gm/Cgg ~1/Lg ) is mainly a function of bias and technology and it sets limit for the gain bandwidth product and noi se performance of a single stage amplifier and charging/discharging time in digital circuits. And Cgg is the sum of the capacitances on MOSFET gate (Cgg =Cgd +Cgs+Cgb). Figure 2-3. NMOS FET model. For a device, higher fmax usually means higher maximum power gain achievable at a given frequency. This is particularly im portant for RF/mm-wave circuits. fmax is highly layout dependent and heavily influenced by th e device loss from parasitics such as Rg, Rs/Rd and the substrate parasitics. Because of this, though fT continues to increase w ith scaling as shown in
38 Figure 2-1, fmax may not necessarily improve with the same pace. fmax can be estimated by as Equation 2-3 : dssggggdmg tgRRCCgR f f )()/(2max (2-3) gds is the small signal conducta nce between drain and source. Based on Pospiezalskis noise model  which matches MOSFET noise performance well at high frequencies, simple estimation for NFmin of CMOS devices is Equation 2-4 : gm tRg f f NF 21min (2-4) and in Equation 2-4 are process/ bias dependent quantities to describe the transistor channel thermal noise source and the ratio of gm/gdo, respectively. When DC bias is fixed, they are basically fixed. It should be noted that even if we use short transistor fingers to minimize the gate resistance Rg, the minimum noise figure is still bounded by the NQS resistance Rnqs seen at the gate, which is about 1/5gm . Observing Equation 2-3 and Equation 2-4, it can be seen that in order to squeeze the best performance out of the device, firstly th e device should be biased for highest gm to achieve peak fT; Second, the device layout parasitics, such as Rg, Rs, Cgd and Cgg, and the drain output conductance gds should be minimized (maximize the output resistance ro). Based on these, a multi-finger layout for the NMOS FET is optimi zed in order to improve its high frequency characteristics and reduce unwanted c oupling, as shown in Figure 2-4. The multi-finger layout with double-sided gate contacts is used to minimize Rg (especially those from contacts and poly-gates), Rs/Rd, Rsubs (substrate resistances of drain-to-body and source to body junctions) and reduce Cdb, Csb by junction sharing. The minimum channel length
39 is used to obtain the highest fT and fmax and the finger width is chosen to be 0.9 m as a good compromise between low Rg and excessive number of finge rs. As explained before, Rg is ultimately bounded by Rnqs and will not improve much with finger width below 0.9 m. On the other hand, too many fingers will increase extrinsic parasitics and lower the power gain. The mobility dependence on finger width mentioned before may come into play as well. The spacing between poly-gates and drain/ source contacts are ~0.1 m larger than minimum allowed by design rule to reduce the fringing capacitances between gate and dr ain/source. Metal1 and metal2 layers are used for the drain and source connections with in the transistor ac tive region. The gate, drain and source connections to upper metal layers are made ~1.5 m away from the active region to lower metal-to-metal sidewall capacit ances. Metal1-sbustrate capacitances can be further reduced by cutting unnecessary metal1 area outside of active region and using a higher metal layer for connection there. When ~18 m wide device with the above layout st yle is used for common source (CS) amplifiers, the small signal resistance looking into the drain node of the device is only ~ 200 from simulation, which limits the maximum power gain of the amplifier. To increase this resistance and hence the gain, just one row of s ubstrate contacts is placed along the bottom of the active area instead of using a ring of contacts surrounding the active area as traditionally done. Lastly, the threshold voltage Vth is ~0.6 V for an NMOS FET in the 65-nm process and the nominal supply voltage Vdd is ~1.2 V. Since the linearity performance of the device is about proportional to the ove rdrive voltage (Vgs Vth) on its gate when it is used in a common source configuration. The relatively high Vth and low Vdd will degrade the linearity performance of circuit.
40 Substrate Contacts Double Gate Contacts Gate Terminal Drain Terminal Source Terminal Figure 2-4. Optimized NM OS FET layout for higher fmax and ro in a TI 65-nm digital bulk CMOS process 2.2.2 Passive Devices in TI 65-nm Digital Bulk CMOS Process As said before, quality factors of on-chip passi ve elements such as inductors and capacitors are critical for RF/mm-wave tuned circuits In silicon mm-wave circuit design ,,, people have been using the planar distributive passives such as transmission lines (t-line), branch-line couplers etc., as was done in traditional microwav e/mm-wave III-V circuit design. These have advantages over lumped pa ssives such as having a more controlled ground return path, robustness against variations, and model scalability. Howeve r, distributive passives occupy a much larger area than lumped passive s in most cases even in the mm-wave band. Larger area means higher cost in IC design, so in the proposed work, lumped passive elements are used to save chip area, mimicking low GHz RF-IC designs in silicon. In some cases where long interconnects are necessary (the distance comp arable to wavelength) and distributive effects
41 may have to be accounted for, e.g., the long differential lines delivering mm-wave LO signal from a frequency synthesi zer to a down-converter. Next the design and modeling of capacitor, induc tor, transformer, and signal pad will be reviewed. As will be seen, and doublemodels  widely employed for low frequency passive modeling remain usable at mm-wave frequencies due to the relative small dimensions of these components. Due to the difficulty of obtaining accurate de-embedded data from measurements of small devices needed for mm -wave applications, much of passive device modeling works here rely on the mode rn electromagnetics (EM) simulations. 188.8.131.52 Metal-oxide-metal capacitor Metal capacitor (cap ) has much higher quality factor than that of MOS capacitor at high frequencies so they are used extensively in low GHz RF-IC design where specialized metalinsulator-metal (MIM) capacito r is not available. The multi-finger comb capacitor is a high density metal capacitor [ 58], but its long, narrow and thin metal fingers in the process raises concern over the relatively high resistance and in ductance that may accompany the finger comb capacitor. This lowers both the quality fact or (Q) and self-resonant frequency (SRF). Fortunately, with the th in inter-level dielectric (SiO2) layers among the lower 5 metal layers in the process, a relatively high de nsity metal-oxide-metal (MOM) capacitor can be formed by overlapping metal plates on different metal layers (F igure 2-5 (A)). In our design, MOM capacitors are mostly made into a square shape to minimize parasitic resistance and inductance. The value of the de sired capacitance between metal pl ates is mainly determined by the area of the plates while the parasitic cap acitor between the metals and the substrate/ground shield relate to both the area of the metal plates and the distance from the metal to the substrate. There is a tradeoff here: smalle r area capacitor can be achieved by using more metal plates, but
42 the parasitic capacitor may increas e due to the narrowing of the gap between the lowest metal plate and substrate. This increases the capacitive loading of circuits when the MOM capacitor is in series with a signal path. A B Figure 2-5. Metal-oxide-metal capacitor. A) Top and side view of a metal3-metal5 MOM capacitor. B) Model for MOM capacitor. With the help of simple MATLAB programs, it is decided that for MOM capacitor used in shunt configuration, metal1 to metal5 should be used to increase capacitance density and reduce capacitor area, since the parasitic capacitor there is not a concern. For MOM capacitor in series, choosing metal3 to metal5 for the plates minimize s the parasitic capacitan ces in percentage of the designed capacitor value. Metal1 ground shie ld plane and HIRES re gion could be added below the series capacitor to isolate the parasi tic capacitor from the substrate to reduce loss and coupling.
43 A simple model is used for the MOM capacitor, as shown in Figure 2-5 (B). The inductance in the model is to account for the inductive parasitics of the capacitor connections and metal plates; HFSS  simulati ons for a variety of capacitor va lues are done and compared with the results from the MATLAB programs to confirm the validity of the MATLAB estimations (Table 2-1). Table 2-1. Capacitances of a series of metal-oxide-metal cap formed by metal 3,4 and 5 from simple MATLAB program and HFSS simulations at ~70 GHz. Ls ~ 3 pH is assumed. Capacitances from MATLAB (fF) Capacitances from HFSS w/o considering Ls (fF) Capacitances from HFSS subtracting Ls effect (fF) 11 10.6 10.5 39 39 38 165 180 165 365 440 367 650 992 662 As introduced earlier, to fulfill the metal density requirement, floating dummy metals are filled around and below the capacitor. They increas e the parasitic capacitances and loss. Because of this, dummy metal block layers are added on me tal 3-5 to make sure the dummies are at least 3 m away from the capacitor metal plates. At the same time, narrow slots on metal plates are added to satisfy the metal slot rules, which do not expect to change the capacitor capacitance much due to the fringing effect of the slots from MATLAB estimation. 184.108.40.206 Metal-oxide-semiconductor capacitor MOS capacitors are formed between polysilic on-gates (poly-gates) and n-well. The extremely thin gate oxide layer gives a MO S capacitor very high capacitance density in accumulation mode. Hence, a MOS capacitor occupies a much smaller area compared to a metal capacitor. However, the quality f actor of MOS capacitor is lower than metal capacitor. With a multi-finger layout and shunting multiple metal laye rs for terminal connections, the resistances from poly gates, n-well, n-well contacts and a ll interconnects are reduced but parasitic junction
44 capacitor and the substrate resistive network intr oduce significant loss. Besides, a MOS capacitor is essentially a varactor, so its capacitance value changes with signal swing. This causes nonlinearity and distortion of signal. The above two disadvantag es limit the use of large MOS capacitors to large bypas s capacitor on voltage supply nodes in mm-wave circuit designs. In this way, the n-well terminal of MO S capacitor will be connected to ground plane so the junction parasitics can be neglected. In this case, MOS capacitor can be modeled simply as one capacitor in series with a resistor to the on-chip ground. A B Figure 2-6. Varactor layouts: A) Top view of the unit-cell. B) Side view of the unit-cell. A small MOS capacitor (varactor) can be used for frequency tuning, such as those in voltage-controlled oscill ators (VCOs). As an example, an accumulation mode MOS varactor
45 , which has estimated Q of ~5 at 94 GHz and Cmax to Cmin ratio of ~3 is used in a 94-GHz VCO . The unit-cell layou t of varactor in  is shown in Figure 2-6 where each polysilicon gate finger (0.64 m X 0.18 m) is contacted from two sides w ith 4 contacts per side to lower the contact resistance. The n-well contact to polysilicon spacing is set to 0.4 m to lower the parasitic capacitance between the two. Usually, the poly-gates terminal of the varactor is connected to the signal path to avoid th e n-well-to-substrate junction parasitics. 220.127.116.11 Spiral inductor Spiral inductors have been used widely in low GHz RF-IC design. At mm-wave frequencies above 50 GHz, since th e inductance needed is usually small (<250 pH), the SRF is expected to extend to higher than 100 GHz with car eful design. This provid es the possibilities of using spiral inductors at mm-wave designs. Nevertheless, quality factor (Q) of inductors implemented in the TI 65-nm CMOS process fo r mm-wave designs are found to be much lower than expectation. The main reasons for this may be the relatively lossy signal path, close proximity of metal to substrate and floating dummy fills, as will be discussed in following sections. Traditionally, the Q of spiral inductor is e nhanced with shunting multiple metals for the spiral traces, and adding a dens e poly-silicon patterned ground shield (PGS) under the spiral . However, neither option is vi able in the process. First, only one thick copper metal (metal6) is available that does not require wide width, while the other lower metal layers (metal1-metal5) are too thin and their distances to the substrate are too small to be helpful by shunting metal layers or adding a PGS. Second, the poly-silicon slabs constituti ng the PGS are not silicided, and the spacing between the slabs ha s to be enlarged to prevent excessive capacitive coupling and eddy current on the PGS. Thus the PGS is rather lossy itself.
46 An alternative is to place only HIRES layer and no shields beneath the spiral traces made of metal6. In this way, the capacitive coupling path from the inductor traces to the substrate sees higher impedance and the magnetically coupled current in the substrate is also reduced. A series of HFSS simulations have proven that the qual ity factors of such sp iral inductors (without floating dummies) could be above 13 from 50 to 80 GHz. In comparison, spiral inductor with a poly-silicon PGS or a metal1 PGS has quality fa ctors below 10 in the same frequency range. Also, from HFSS and Fasthenry [6 3] simulations, it is found that an inductor trace width of ~1.6 m (5~6 times of the skin depths of 60-80 GHz ) combined with a trace spacing of ~2.5 m and a hollowness factor  of > 1/3 would provide the highes t inductance and Q. These avoid excessive parasitic capac itance between the traces and substrat e, reduce the side-wall capacitance among the traces, and alleviate the proximity effect. Unfortunately, the floating dummy fills added by the process will degrade the spiral inductor performance. To minimize the excessive loading and undesired coupling effects from these dummy structures to the inductors (and other mm-wave signal paths) while satisfying the metal density rules, designed dummy blocks and metal fill patterns have to be included as part of the design, floor-plan and layout considerations which reduces the area advantage of lumped passives. Custom metal1-6 patterned meta l fills are added inside and ou tside of the spiral inductor loop as shown in Figure 2-7 (A). The ultimate goal is to mitigate the loading effects of the dummy metals on the inductor and reduce parasitic capacitances of the inductor. The principles of adding these dummy pa tterns are: avoid dummy metal dir ectly below the inductor trace, maximize the distance between the induc tor trace and the dummy fills (~2.5 m), use small area dummy fill elements to prevent possible e ddy current flow in the dummy element (~2 m x 0.8
47 m) and reduce the parasitic capacitance to s ubstrate, and maximize the distance among dummy elements (~0.2 m) to improve their isolation. Figure 2-7 (B) plots the same spiral inductor with uniform metal1-6 patterned dum my fills added by the foundry automatically if no custom patterned fills are used. A B Figure 2-7. Spiral induc tor. A) Top view of a rectangular spiral inductor with customized dummy metal pattern fills. B) Top view of a rectangul ar spiral i nductor with uniformly distributed dummy metal fills emulating dummy metals added by the foundry. C) Model for the spiral inductor.
48 C Figure 2-7. Continued HFSS simulations showed that comparing sp iral inductor with the optimized dummy patterned fills described above a nd that with rela tively random dummy patterned fills, Qs are ~9 and ~6 respectively from 60-80 GHz. Such large de gradation of Q compared to that of ~ 13 for the case without dummy metals, may be partially due to the concentrated magnetic field in the center of the spirals causing eddy current loss on the dense dummy metals there. In our circuit design, a doublemodel  is used for induct ors and short in terconnects to account for the distributive effect at mm-wave frequencies (Figur e 2-7 (C)). The values of the doublemodel parameters can be curve-fitted to the simulated data from HFSS (Figure 2-8). 18.104.22.168 Line inductor In many cases, mm-wave tuned circuits only ne ed inductors with va lues of <150 pH, a short metal line could be used for that purpose. These line inductors can be implemented with or without metal ground shield as shown in Figure 29. Compare the three cases in HFSS (Figure 210), the one without a ground shield but with HIRES region under the line has the highest inductance and Q for the same length. The reasons for this are that there is no metal ground plane below the line, so the main return current path is further away from the forward current path on the line, hence the inductance per unit length is raised. Besides, the capacitiv e loading to the line
49 without ground shield is less than the other two cases, which extends SRF and increases inductance. A 1.6m wide and 150m long has an inductance of ~210 pH and Q of ~13 from 50 to 80 GHz for the case without a ground shield, in contrast to an inductance of ~190 pH and Q of ~9 for the case of with sa w-tooth shaped shield, and an inductance of ~60 pH and Q of ~7 for the case of with a slotted ground plane. 0 50 100 150 200 250 300 6065707580Frequency (GHz)Inductance (pH) Custom Dummy Met a Fills Random Dummy Met a Fills No Dummy Metal Fills A 0 5 10 15 6065707580 Frequency (GHz)Quality Factor Custom Dummy Metal Fills Random Dummy Metal Fills No Dummy Metal Fills B Figure 2-8. HFSS simulated result s for a spiral inductor. A) Extr acted inductances of structures in Figures 2-7 and that without the metal fills. B) Extracted Q s of the structures.
50 A B C Figure 2-9. Various line inducto r structures A) Line inductor without a ground shield. B) Line inductor with a metal1 tooth-shaped ground shield. C) Line inductor with a slotted metal1 ground plane underneath (microstrip transmission line). D) Line inductor with floating metal1 strips undern eath as a shield . E) Line inductor without a ground shield and with custom dummy metal fills on two sides.
51 D E Figure 2-9. Continued It is much easier to include dummy block a nd custom metal fills on the two sides of line inductors compared to spiral i nductors. As long as the floor plan and chip area permits, the metal dummy fills should be placed as far away from the line inductors as possible. Moreover, from HFSS simulations, if dummies are put ~3 m away from the line inducto r in Figure 2-9 (A), then inductance only reduces by ~2% and inductor Q onl y degrades by ~2 compared to the case of without dummy fills at all. This is probably due to the circling magnetic fields of line inductors are concentrated in the proximity of the line a nd hence not influenced much by the dummy fills
52 on the two sides. Similar to spiral inductors, a doublemodel is used to model the line inductors. As a final note, with a slotted ground plane under the line (Figure 2-9 (C)), it actually becomes a microstrip t-line. Through HFSS simulati ons, we can see that a mi crostrip t-line (and A 0 5 10 15 20 50607080 Frequency (GHz) Q ua li ty F actor Line ind. w/o gnd shield Line ind. w/ toothshaped gnd shield Line ind. w/ complete gnd shield Line ind. w/ floating shield Line ind. w/o gnd shield & w/ custom dummy metal fills B Figure 2-10. HFSS simulate d results for various line inductor structures (40 m long and 1.6 m wide metal6 line) in Figure 2-9.A) Extracted inductances. B) Extracted Qs.
53 co-planar waveguide (CPW) t-line) does not have good inductive Q and needs a much longer length to reach the same inductance as the one w ithout a ground plane. So t-line is not used to form resonant tank and matching networks. 22.214.171.124 Transformer On-chip transformers can be implemented w ith two inductive windings sharing the same normal plane for magnetic fields. The two windings usually called primary and secondary, can be built on the same metal layer or not. Tran sformer can conveniently provide DC isolation, biasing, and reactance for impedance matching between two RF stages It is also often used as a single-to-differential converter (balun). Such an example is shown in Figure 2-11 (A), which includes a top view diagram of a transformer constructed in metal6 of the TI 65-nm process with two single loop windings and its lumped model ba sed on 2-port Z-paramete r fitting. If port 1 (primary winding) of the structure is connected to a single-ended source, then the two branches of port 2 (secondary winding) terminals will ha ve out-of-phase signals coupled from port 1. Here, we call it a loop balun and it was used in a 77-GHz down-conversion mixer design. When the two ports are matched to 50in the range of 50 to 60 GHz, the loop balun shows ~3 dB power loss from port 1 to port 2 in HFSS simula tions. To lower that loss, both the Q of two windings and magnetic coupling between them should be maxi mized while the capacitive coupling between the two should be minimi zed. The loop balun has non-trivial magnetic couplings between the two differen tial output branches, which is reflected in its model with the coupling coefficient K32 in Figure 2-11 (B). This causes an inherent asymmetry at the output port and common mode signal will be generated from this, which should be minimized. An alternative balun to alle viate this issue can be built by putti ng two straight lines close to each other, which is called line balun here (Figure 2-11 (C)). The two differential output terminals
54 A B C Figure 2-11. Baluns: A) The loop balun in metal6 (not show ing custom dummy metals). B) The lumped model for the loop balun. C) The lin e balun in metal6 (not showing custom dummy metals). then can be far away from each other and their mutual coupling is negligible. The line balun has less than 4 dB power loss from measurements in 65-80 GHz, it occupies smaller area, makes floor-planning much easier and simplifies the fitt ing of the lumped mode l in Figure 2-11 (B),
55 where the K32 is negligible now. Simulations of two-por t S-parameters for a loop balun tuned at 50 GHz and a line balun tuned at 80 GHz are done in HFSS and the re sults show that the baluns have ~3 dB power loss in respec tive frequency range (Figure 2-12). -24 -18 -12 -6 0 354555657585 Frequency (GHz)S11, S12 (dB) Loop Balun Tuned to 50 GHz Line Balun Tuned to 80 GHz Figure 2-12. Simulated |S11| and |S12| of a loop balun tuned to 50 GHz and a line balun tuned to 80 GHz from HFSS two-port simulations. 126.96.36.199 Signal pad Signal pads are essential for connecting circuits on-chip to the outside world except in the case where communications can be achieved with on-chip antennas  or transformers. They also play critical roles to support probe landing for on-die measurements. Usually, the simple model of series RC to ground is used for signal pad (Figure 2-13). In the low GHz RF-IC design, much attention is not being paid to signal pad because the pad parasitics give relatively high impedances and do not infl uence circuit performance much. However, at mm-wave frequencies, the parasitics of signal pad show larger effect on circuit behavior, e.g., in the TI 65-nm process, a rectangular signal pad with the size of ~ 50 m X 50 m is found in measurements to have ~80-fF cap acitance from the top al uminum metal of the pad to the metal1 on-chip ground shield which translates to less than 30 impedance at 67 GHz. This really complicates the design of mm-wave input ma tching and degrades the input
56 matching Q. The 80-fF pad capacitance is much la rger than our original estimation of ~40 fF based on the pad area, probably due to the dummy metals surrounding the pad. To reduce this capacitance, dummy blocks are adde d on each metal layers of the pa d so that all the dummies are ~5 m away from the pad. HIRES layer is also pl aced below the pad to reduce the substrate loss as well. In order to mitigate the pad parasitics issue in high frequency circuits, the pad area may have to be reduced further, or pad shape may need to be changed from a rectangle to a diamond . But ultimately, the pad area has a lower limit for bonding or probe landing. Tuning the pad capacitance out with a shunt t-line  seems to be a viable solu tion, but this method will have to be carefully evaluated with EM tool to prov e its effectiveness in the TI 65-nm process. Figure 2-13. Signal pad structure and its model 2.3 Summary In this chapter, general trends and issues of nano-scale digital CMOS process and mmwave circuit design directions are discussed. Simple, robust circuit topologies with lumped passive tuning elements should be explored. Next, the features of TI 65-nm digital bulk CMOS process are introduced, and the modeling and optimization of th e NMOS transistor in this process are described. Finally, th e design and characteristics of a se ries of passive devices in TI
57 process are explained with simulated and measured results. These devices will be used in the mm-wave circuit designs for this thesis research.
58 CHAPTER 3 CRITICAL CICRUIT BLOCKS OF MILLI METERW AVE RECEIVER FRONT-END 3.1 Introduction Usually, a wireless receiver front-end consists of multiple circuit blocks cascaded together. Along the receiver chain, received signal on a high frequency carrier will be filtered, amplified and mixed down to desirable low frequency. Extr a noise generated by the circuits will be added to the signal and hence the output signal-to-noise ratio (SNR) will be lowe r than the input SNR. Such SNR degradation due to circuits is an inherent property, and it can be characterized as the noise factor F (noise fi gure NF = 10log(F)). ,outout ininNS NS F (3-1) In a complete system consisting of n-stages, if the individual noise factor Fn and available power gain Gn are known, the overall noise factor can be calculated by Friis equation (Equation 3-2). 1 21 21 3 1 2 11 1 1 n n sysGGG F GG F G F FF (3-2) Equation 3-2 shows the overall sy stem noise factor is mainly determined by the first few stages if they have sufficient gain to suppress the noise contributions fr om the following stages. Based on the discussion in Chapter 2 of the TI 65-nm low-cost low-leakage bulk CMOS process and active/passive devices, we will de sign and implement several fundamental circuit blocks for mm-wave receiver front-end using the process to explore the feasibility and performance of these blocks in the low-leakage bu lk CMOS. This serves as a first step towards building more complete low-cost mm-wave transcei vers and systems for pract ical applications as suggested in Chapter 1.
59 3.2 Tuned Amplifier A millimeter-wave tuned amplifier is one of the most important blocks in the receiver front-end and it can be used as the first amplif ying block for the incoming weak signal. As can be seen in Equation 3-2, its noise factor directly adds to the ov erall system noise factor and its power gain plays a pivotal role in suppressing th e noise from the following stages. As will be discussed shortly, an 80-GHz tuned amplifier is designed, and it has relatively moderate power gain and high noise figure compared to the low GHz frequencies counterparts as expected due to the limited device performances at mm-wave freque ncies. Because of this, the amplifier is not referred as a low noise amplifier (LNA). 3.2.1 Design of 80-GHz Tuned Amplifier As shown in Figure 1, an 80-GHz stagger-tuned amplifier is designed a nd it consists of six common source (CS) amplification stages cascaded together. Each NMOS tr ansistor uses a shunt inductor (Ld), a series capacitor (CS) and a series inductor (Lg) for inter-stage matching. The simpler the matching topology, the lower loss the signal experiences from the passives at such high frequencies. The six stages are needed to achieve the desired gain an d bandwidth with some margins since each CS stage has limited power ga in (~2-3 dB) especially with the moderate Q passives. The input matching network of the amp lifier is designed for low noise operation with acceptable power match . The rest of networ ks are conjugately matched for higher power gain, while the output network is optimized for wide band matching. The use of multi-stages also increases the input-output isolation and mitigates the relatively poor isolation of common source amplification stages. The CS topology is chosen in favor of th e widely utilized cascode topology - for its better performances around 80-GHz . The relatively low supply voltage of 1.2 V and relatively high threshold voltage of ~ 0.6-0.7 V for the low leakage transistors leave small voltage headroom. This situation is partic ularly challenging in a cascode stage. Reduced
60 Vdss for both the common-gate (M2) and bottom common source (M1) transistors lower their transconductance (gm) and small signal output resistance (ro). Single stage 80-GHz CS and cascode amplifiers (Figures 3-2 (A) and (B )) are designed and optimized for comparison. Inductor Qs of 12 and capacitor Qs of 10 are assumed. Figure 3-1. Schematic of the 80-GHz 6-stage tuned amplifier. It is found that a cascode stage with the optimum W2/W1 ratio of 2 can have ~1 dB higher gain and ~0.7 dB lower noise figure (NF) than that of W2/W1 of 1 as done in - and . A wider M2 increases the Vds1 of M1 and enhances its gm1 by ~17% and ro1 by ~44%, while gm2 ro2 is also increased by ~18%. Furthermore, inse rting a series peaking inductor  in the cascode device can increase its power gain by ~1 .2 dB and reduce its NF by ~0.4 dB. However, even with all these, the cascode amplifier still shows ~1.3 dB lower gain and ~1 dB higher NF than the one-stage CS amplifie r around 80 GHz, as shown in Figure 3-2 (C). The main reasons for this are the lower gm1 and ro1, and relatively low impedances from middle node parasitic capacitances at mm-wave frequencies in a cascode device.
61 A B -5 0 5 10 70758085 Frequency(GHz)|S21|(dB)5 10 15 20NF(dB) |S | of CS amp. |S | of cascode amp. NF of CS amp NF of cascode amp. 21 21 C Figure 3-2. Comparison between CS and cas code 1-stage am plifier around 80 GHz. A) Schematic of the CS amplifier. B) Schematic of the cascode amplifier with series inductor peaking. C) Simulated power gain and NF for the two 1-stage amplifiers.
62 The drain output resistance of cascode stage a nd hence its power gain is more sensitive to the change of transistor substrate resistances Rsub1, Rsub2 and Rsub3 (Figure 3-3) compared to the CS amplifier. Referring to a simplified NMOS tran sistor model based on 2-3 in Figure 3-3 (A), the equivalent CS device drain output resistance Rd is found to not vary mu ch with the change of Rbody because the CS intrinsic ro ~ 1/gds is comparable to the magnitude of Zdb (the impedance of Rbody-Cdb series combination). In contrast, the cascode equivalent drain output resistance is more sensitive to the change of Rbody and Cdb for the common gate transistor, since its intrinsic drain output resistance Rdcas (~gm2 ro2 (Zd1||Zgs2)+ ro2) is higher than that of a CS stage. This implies that a CS amplifier has less performance variation caused by the transistor body resistance uncertainty, as shown by the simulated re sults in Figures 33(B) and (C). A Figure 3-3. CS and cascode amplifier perform a nce sensitivity with device body/substrate resistance. A) CS amplifier with a simplified NMOS transistor model. B) Simulated power gain and drain output resistance of the CS amp lifier versus transistor body resistance. C) Simulated power gain and drain output resistance of the cascode amplifier versus transistor body resistance.
63 50 100 150 200 250 100200300400500 Rbody( )Rd( )2 3 4 5 6Gain (dB) CS device R 1-Stage CS amp. gain d B 0 200 400 600 800 100200300400500 Rbody1,2( )Rd2( )0 2 4 6 8Gain (dB) Cascode device R 1-Stage Cascode amp. gaind2 C Figure 3-3. Continued A CS stage has worse inputoutput isolation and it can be unstable due to the Cgd, Cds and Cgs coupling . The six stagger-tuned CS stages in the 80-GHz amplifier use the same optimized NMOS transistor layout described in Section II. The tr ansistor width and length are 18 m and ~65 nm for all stages, which provides an input matching network Q of ~ 2.5 for acceptable matching bandwidth and to lerance to matching element vari ations. The transistor gate bias is selected to maximize gm in simulations. This allows use of narrower transistors with reduced parasitic capacitances a nd associated coupling between i nput and output of each stage, and improves the overall stability. In simu lations, a CS amplifier stage exhibits |S12| of less than
64 dB from 60 to 90 GHz. Small source degenera ted inductors ranging from 10 to 40 pH are used to make the amplifier stable. The multi-stage amplifier topology is relatively si mple but the design flow is not as simple. Since so many stages are involved, they interact with each other. Besides, many device parasitics that are neglected at lower frequency desi gn may no longer be negligible at mm-wave frequencies. The analytical desi gn method used in low GHz RF circ uit is almost too complicated to use for a time-limited design, although it coul d potentially lead to better understandings of design parameter tradeoff and optimization. Here we resort to the traditional microwave design method, which simplifies the design process by tr eating devices as a two-port black box. The design flow is described in steps as follows: First, for the stagger tuning, frequency planning among stages is needed to cover the required bandwidth with a relativel y flat gain. Suppose we need to cover the ba ndwidth from 74 to 80 GHz with a gain of 15 dB with some marg ins for tuned gain and frequency variations, we roughly need to design for a gain of 18 dB or mo re with 3 dB bandwidth covering 76 to 86 GHz. Second, based on an initial simulation for gain and 3-dB bandwidth fo r 1and 2-stage CS amplifier at ~80 GHz, about six cascading stages are needed to achieve our design goal. Each stage tuned frequency then is de cided initially as: Stage 1 and St age 6 need to be tuned at 81 GHz to have the input and output matching centered at desired band; Stages 2 and 5 are tuned at 84 GHz with the latter one leaning towards 81 GHz to boost the relatively weak gain at the high end of the band. Besides, due to the finite c oupling between Stage 1 and 2, and between Stage 5 and 6, Stage 2 and 5 tuned at ~84 GHz can help expand input and out put matching bandwidth; Stages 3 and 4 are tuned between 78 GHz and 81 GHz to flatten the overall gain. Since Cgd, Cgs and Cds of transistors provide feedback (coupling) from stage to stage, the above tuned
65 frequencies for each stage will need to be ad justed. The selection of tuned frequency is qualitatively shown in Figure 3-4. Third, select NMOS transistor size by trading off among several factors. Firs of all, the transistor should have adequate maximum stable gain or gm depending on the application without excessively high power consumpti on. And it is desirable to have a moderate first stage input matching network Qin, and hence larger matching components variation tolerances. Also, the transistor size should be kept moderate to avoid degraded terminal Qs and larger coupling among terminals due to increased pa rasitics influences. It should be noted that the selection of transistor size interacts with selection of appropria te inductive and capaciti ve load values for optimum overall Qs of the transistor terminal s. Usually, these passive components have lower self-resonant frequency (SRF) and lower Qs with larger values (The de tails are surely highly dependent on the process). The transi stor finger width is set to ~0.9 m for optimum high frequency performances as discussed in Chapter 2. Figure 3-4. Stagger tuning of six amplification stages
66 Fourth, after the NMOS transistor is sized, it will be connected as shown in Figure 3-5, where the inductors and bypass cap will employ the models introduced in Chapter 2 (It must be emphasized here for the importance of using compone nt models as close to the reality as possible to obtain accurate results). Then optimum source impedance reflection coefficient (GS) can be explored in S-parameter simulation in such a schematic. After the amplifier output resonant frequency is adjusted to the desired one as pl anned earlier, then the input impedance, noise circles, available power gain circ les, stability circles at the desi red frequency are examined with different Ld and Ls values until a Sopt (ZSopt) is found for a good comprise between power gain, noise figure and stability using a smith chart, as shown in Fi gure 3-6. Fortunately, the input impedance seen at the gate is not a strong function of Ld and Ls as a result of choosing a moderate transistor size. In th is way, all six stages are rough ly tuned at desired frequencies independently. Since identical tran sistors are used for all six stag es, their input impedances and S11* do not change significantly even though the st ages are tuned at different frequencies. Figure 3-5. One-stage CS amplifier sc hematic used to tune frequency and find Sopt to obtain a good tradeoff among power gain, no ise figure, and stability. Ccouple is just an ideal large capacitor to f acilitate correct S-parameter simulation.
67 70-90 GHz Optimum Source Impedance (ZSopt) For Low Noise Source Stability Circle (SSC) Available Power Gain Circle (GAC) @ 80 GHz Noise Figure Circle (NC) @ 80 GHz NF=5 dB NF=4 dB ZSopt(~31+j78 @ 80 GHz) GA=5 dB GA=7 dB SSC@70 GHz SSC@80 GHz SSC@90 GHz Figure 3-6. ZSopt, availabe power gain circ les, noise figure circles a nd source stability circles around 80 GHz for low noise matching network design.The stable regions are outside of stability circles. At this point, the input matching to the firs t stage CS amplifier, which includes the pad capacitance and a series Lg as shown in Figure 3-7, can be designed by transforming the 50source impedance to Sopt as close as possible due to the non-ideal parasitics of the matching components. After that, the impe dance reflection coefficient S22 looking into the transistor drain is found around the tuned frequency of the following stage by replacing the Ld with an RF choke, as also shown in Figure 3-7. Then S22 can be transformed close to Sopt for the following stage using the inter-stage matching network shown in the dashed box in Figure 3-7. The following stage is connected to its next stage through the same kind of inte r-stage matching network, which component values can be found approximately from the S22 and Sopt in Figures 3-5 and 3-6.
68 The Sopt point for one stage can be either closer to the optimum noise matching or conjugate power matching. Figure 3-7. Inter-stage ma tching networks schematic When stages are cascaded, the input impedance of an individual stage must be checked again after they are connected to the following stage through the inter-stage matching network, and the input matching network for the particular stage may have to be modified according to the change of its input impedance. However, the modification to the input-matching network will also influence the output impedance of the st age. Fortunately, the coupling between the device gate and drain is not strong, and the inter-stag e matching network tuned at different frequencies helps the isolation among different stages, thus the modifications from such iterations are not significant and a large number of iterations is not needed be fore a good result is reached. Neutralization techniques to s uppress the input-output feedback, such as adding an inductor
69 between the device gate and drain to tune out the Cgd , are not used to reduce the complexity of the multi-stage design. Finally, the gain, 3-dB bandwidth (BW), NF input and output matching, stability and component variation tolerances of the whole amplifier must be examined. Adjustments for matching component values found in previous st eps may be needed to achieve better overall simulated performance. After the somewhat tedious design procedure, a reasonable design is achieved. However, non-ideal ground plane and voltage supply network, and lossy s ilicon substrate coupling are additional factors that need to be considered at mm-wave frequencies. These will be discussed in the following sections. 3.2.2 Non-Ideal On-Chip Ground and Supply Network Usually, silicon integrated circ uits, unlike traditional III-V microwave designs where there is a thick backplane metal and lo w-impedance large vias connecting top side metal layers to the backplane, must have explicit metal traces/pla nes drawn as on-chip signal reference plane and current return path. And we call it on-chip ground. When signal current flows (we called it forward current here) betwee n an external signal source /terminator and on-chip ci rcuits through a package or a probe, return current in the opposite direction of forward current has to flow through the on-chip ground back to the reference point of the external source/terminator Figure 3-8 illustrates such a scenario, which includes non-ideal parasitics on bo th DC and RF current paths in the on-chip/off-chip system. For low GHz RF-IC design, the on-chip ground t ypically is viewed as an almost ideal ground with very little impedances so that for all on-chip compone nts, they can be modeled as having the same reference points. Ground bounce may still occur because of the non-negligible package impedance from the chip ground to the external source/terminator ground. This issue
70 could be mitigated by connecting the on-chip ground to the external ground with multiple downbonds or multiple balls in flip-chip technology. Nonetheless, for mm-wave circuits in a digital CMOS process such as the TI 65-nm process, th e relatively thin metals (as ground planes), the high operating frequencies and the skin effect al l makes the parasitic inductance and resistance non-negligible for the return currents. Figure 3-8. Parasitics in the current paths including on-chip/off-chip circuits Moreover, since lumped passives such as spiral inductors or lines induc tors are used in our design, if the ground plane is completely cl osed around the inductive passives, then their inductances and Qs degrade dramatically due to more signal return current and eddy current induced on the ground plane, as shown by HFSS simula tions in Sections 188.8.131.52 and 184.108.40.206. So the ground plane surrounding theses components ha s to leave a gap open in the on-chip ground. This increases the return path impedances.
71 In addition, distributive natu re of the on-chip ground comp licates matters more. The mmwave circuit layouts are not as compact as what we have hoped for. The whole chip sizes and hence the on-chip ground plane sizes in many cases are more than or comparable to a quarter of the signal wavelength. The relatively large chip area is mainly attributed to designed metal dummy fills associated with most passive compone nts and the fact that due to low power gain achievable per stage, a multi-stage amplifier must be used to obtai n the desired gain. In a word, the expectation of an ideal on-chip ground for on-chip components at mm-wave frequencies is not realistic. Besides, the para sitic impedances of the non-ideal ground plane could provide paths for undesirable coup ling among stages and introduce di stortion to the signal. When modeling components, attention shou ld be paid to the ground return path parasitics, especially for lumped elements. Because their ground return paths are highly layout dependent and not very clear sometimes. (In contrast, transmission lines have well-defined groun d return paths, which are taken into consideration expl icitly in the modeling process.) To alleviate this non-ideal ground issue, the on-chip ground should be made as low impedance as possible, and the non-ideal ground eff ect should be modeled and incorporated with the design in simulation. For the on-chip ground layout, a ground unit-cell is drawn so that it can be replicated periodically and form a ground pl ane while both the metal slot ru le and metal density rule are satisfied. After the main on-chip components ar e connected in the layou t, the ground plane can be placed wherever there are spaces within th e chip boundary, and ground terminals of some components such as bypass cap will be connected to the ground plane. Mu ltiple-metal-stacked ground unit-cells should be used as long as they do not create undesire d capacitive loading to high frequency signal lines of the circuits. On -chip ground floor plan and layout process are
72 time-consuming, however, they are vitally im portant for proper operation of RF/mm-wave circuits. For modeling and verification of on-chip gr ound parasitics effects, first, the grounding nodes more sensitive to the non-ideal ground effect than others need to be identified. In general, the signal nodes with a low impedance path to the on-chip ground are sensitive. For the 80-GHz tuned amplifier, the groundi ng nodes of the bypass caps, Ls, and series inductor or cap that has relative large parasitic caps to ground are the sensitive nodes. Next, the whole slotted ground plane can be reproduced in Fast henry and analyzed as an nport network, generating n-port Z parameters that can be directly used in frequency domain simulation. For transient simula tions, RLC equivalent networks among the n-nodes have to be extracted out of the Z-parameters . This method provides a relatively accurate impedance prediction for the ground plane, but the size of the n-port Z paramete rs could be huge for circuits with a large number of sensit ive grounding nodes. The computati on of the RLC extraction is not trivial either, esp ecially when distributive effect has to be accounted. To save time, on-chip ground impedance network among sensitive nodes is roughly modeled using the equivalent RLC model for a short segment of on-chip ground plane with a finite width. The parasitic capacitance in the equivalent network can be regarded as the capacitance between the on-chip ground plane and nearest large area earth ground such as the metal chunk of a probe station, which s hould be small and may be neglected. Then multiple pre-computed RL impedances that are shunted and cascaded between every two nodes connect all sensitive nodes. The distan ce and width of the ground path between two sensitive nodes determine the number of cascad ing and shunting RL sections. Notice such a ground impedance network needs to be connected to the on-chip reference points of the external
73 source/terminator, which explicitly are the onchip grounding pads. Figure 3-9 illustrates the tuned amplifier schematic with non-ideal ground parasitics. Figure 3-9. Simplified amplifier schematic with on-chip ground parasitics network (in dashed box) connecting the sensitive nodes in the tuned amplifier. Although only the impedances of the on-chip ground are discussed so far, the voltage supply network also suffers similar parasitic effects. Large on-chip bypass caps on supply nodes are used to decouple the circu its from the supply network para sitics. However, the non-ideal ground and the use of imperfect/inadequate bypass cap s degrade the effectiveness of these caps. They could cause coupling through supply network and even oscillations. In the design, wide top metal lines (metal6) are used for Vdd routing to reduce parasitic resistances, and ground planes are drawn as close to the Vdd traces as possible to suppress the parasitic inductances. Besides, Vdd supplies for different on-chip circuit blocks s hould be separated and d ecoupled better through larger external bypass caps on pack age or probes. Shared common Vdd paths should be avoided.
74 The simulated results (Figure 3-10 (A)) show th at these non-ideal effect s can degrade NF and gain by as much as 3 dB and 6 dB respectively, in which the worst case results are obtained assuming a section of the slotted ground plane in metal1 with ~80 m length and ~60 m width connects every two ground sensitive nodes of in Figure 3-10 (B) (The section can be with other dimensions and shapes depending on the actual layout). Most return cu rrents around 80 GHz in the circuit should flow within such sections to the RF gr ound pads. The parasitic R (~0.25 ) and L (~25 pH) at around 80 GHz from this me tal1-6 stacked ground strip section can be estimated using Fasthenry (Figure 3-10 (B)). Figure 3-10 (C) shows an ea rlier version of the 80GHz tuned amplifier targeted wi der 3-dB power gain, however, its estimated ground parasitics are much higher because only metal1 is used for on-chip ground plane there. 0 10 20 30 7275788184 Frequency(GHz)|S21|(dB)5 10 15 20NF(dB) |S | w/ ideal gnd |S | w/ non-ideal gnd NF w/ ideal gnd NF w/ non-ideal gnd 21 21 A Figure 3-10. Esitmation of ground parasitics in fluence on amplifier performance A) The power gain and NF of the 80-GHz tuned amplif ier with and without non-ideal ground and supply parasitics.B) A section of slotted metal1-6 ground plane connecting Stage 1 and 2 amplifier sensitive gr ounding nodes drawn in Fasthenry and die photo of the corresponding circuit are show n as an example of estima ting the ground parasitics. C) The power gain and NF of another version of 80-GHz tuned amplifier for wider 3-dB bandwidth with and without non-ideal ground and supply parasitics (only metal1 layer is used for on-chip ground in this version).
75 B -10 0 10 20 30 70758085 Frequency(GHz)S21(dB)5 15 25 35 45NF(dB) S w/ ideal gnd S w/o ideal gnd (worst case) NF w/ ideal gnd NF w/o ideal gnd(worst case) 21 21 C Figure 3-10. Continued Looking at a bigger picture, as already shown in Figure 3-8, the package or test probe parasitics up to the DC or RF sources should be included in the simulation. For simplicity, to include the package/board effect in the circu it design, transmission lines matched to the RF
76 sources are connected to the package bond pads, and then to the chip through bond wires or flipchip balls, while DC source paths are mode led as multi-section lumped RLC circuit. At the circuit topology level, di fferential circuits should be able to alleviate the non-ideal ground and supply effects, which ar e desirable features for highly integrated circuits. However, the effectiveness of differential circuits in th is respect depends on their layout symmetry, which should be carefully controlled. As can be seen, since at least one design-la yout needs to be done before the non-ideal effects from on-chip ground and supply net coul d be evaluated and necessary modifications could be made accordingly, finishing the design and implementation of mm-wave circuits in one round is very challenging. 3.2.3 Substrate Coupling Based on , , a semi-conducting silicon substrate can be modeled as a complex RC network as shown in Figure 3-11 (A) and (B). Th e substrate shows a trend of being with lower impedance at higher frequencies, which cause s not only larger signal loss but also more unwanted mutual interference for on-chip components. In TI 65-nm process, the proximity of the metal layers to the substrate makes matters worse by having potentially larger capacitive coupling. When operating frequencies are increased, the effectiveness of using a P-sub guard ring (and/or n-well guard ring) (Figure 3-8 (B)) to reduce substrate resist ances and isolate individual de vice or circuit block on-chip decreases quickly . Besides, noticing the guard ring structures rely on an ideal AC ground to have good isolation, it is unclear whether the guard rings bring harm s or benefits to the isolation among devices at the presence of the non-id eal ground effect, espe cially at mm-wave frequencies. More studies have to be conducted to clarify this and find effective isolation means for mm-wave designs in TI 65-nm process. However, putting HIR ES regions wherever possible
77 A B Figure 3-11. Substrate networ k. A) The semi-conductive silicon substrate is modeled by a RC network, which provides coupling paths among on-chip circuits. B) P+ contacts to the substrate (guard ring) can be added in-b etween circuit blocks to isolate them. on the die should help reducing th e surface currents in the substrate, and hence its loss and coupling effect. This is a practice adopted in our mm-wave designs. 3.2.4 Measurements of 80-GHz Tuned Amplifier A micrograph of the earlier version of tune d amplifier is shown in Figure 3-12. The amplifier chip size including bond pads is 0.64 X 0.38 mm2.
78 Figure 3-12. Die photo of the 80GHz 6-stage tuned amplifier. The S-parameters of amplif ier are measured on-die with GGB W-band (75-110 GHz) GSG probes and Agilent power network analyzer (PNA) E8361A usi ng SOLT calibration . Figure 3-13 plots the measurement set-up and Figure 3-14 shows the S-para meters from 73 to 83 GHz at Vdd of 1.2 V and bias current of 27 mA. A maximum gain of 12 dB is observed at 79 GHz, and the input and output return losses are better th an 8 and 10 dB respectively within the 3-dB bandwidth between 75 and 81 GHz. Figure 3-13. S-parameter measurement set-up for the 80-GHz tuned amplifier
79 Figure 3-14. Measured S-parameters with 1.2 V supply and drained current of 27 mA. |S12| is below -49 dB within the 3-dB band, indicating excellent reverse isolation. The measurement of amplifier noise figure is done in two steps with the set-ups shown in Figure 3-15. First, the noise figure meter is calibrated. The W-band noise source from NoiseComm is connected to a se ries of components such as Wband isolator, waveguide-cable adapter (Anritsu), coax cable (Gore), W-band down-conversion mi xer (SpaceK) and an external MITEQ IF amplifier (suppressi ng the noise contribution from the noise figure meter) before finally reaching the HP noise figure meter 8970B and test set 8971C configured in the high frequency band mixing mode. The W-band excess thermal noise from the source is filtered by the W-band isolator and waveguide section, then down-converted to the IF band of 13~18 GHz, amplified by the external IF amplifier, then down-converted again to below 1.6 GHz range and measured by the noise figure meter. The ~60 GHz LO signal for the external W-ba nd down-conversion mixer is provided by an HP 83680A signal source cascaded with a NEXTEC power amplifier and a frequency tripler.
80 The ~16-GHz IF LO signal for the second downconversion inside the noise figure meter is provided by an HP 8341A source contro lled by the noise figure meter. Figure 3-15. Noise figure measurement set-up for the 80-GHz tuned amplifier After calibration, the on-die tuned amplifier with two W-band GSG probes landed on its input and output pads, is insert ed into the calibrated measuremen t set-up before the external Wband mixer. The loss of the additional W-band cable and probes are measured separately and used in the amplifier gain and NF data de-embedding. Figure 3-16 shows that the measured NF of the amplifier is less than 10.5 dB within the 3dB bandwidth, and the NF at 79 GHz is 9 dB. It al so shows the NF is less than 9.5 dB within the 3-dB bandwidth and reaches a mi nimum of 8 dB at 80 GHz at Vdd of 1.5 V and Id of 33 mA.
81 Figure 3-16. Measured power ga in and noise figure with 1.2 V supply, bias current 27 mA, and those with 1.5 V supply, bias current 33 mA The differences between the measured and simu lated amplifier gain and noise figure are ~5 and ~2 dB respectively. These ar e attributed to the modeling in accuracies in both the active and passive devices and the variations of component values at mm-wave frequencies. More accurate models may be obtained from finer EM simulations and test structures measurements. Also, adding tuning elements such as varactors and vari able inductors to the amplifier may be able to alleviate the variation problems. Figures 3-17 (A) and (B) show the block diagrams of measurement set-ups for the amplifier input 1-dB compression point (IP1dB) and the input IP3 (IIP3) respectively. As can be seen, two W-band signal sources are needed in the measurements. Figur e 3-18 illustrates how these two signals are generated. W-band signa l source1 is provided by an Agilent N5260A Wband extension module (essentially a series of fr equency multipliers and amplifiers working at 67 to 110 GHz) with its output frequency cont rolled by Agilent PNA E8361A. The W-band single tone output power from the extension module can be adjusted by the pre-installed attenuator on the module and the output power value at the end of the ext. module waveguide section can be monitored by the PNA. W-band signal source2 is provided by another N5260A
82 extension module with it s input connected to an Agilent signal generator E8254A. The signal generator can control the modul e output frequency and power. A B Figure 3-17. Linearity m easurement set-ups: A) IP1dB measurement set-up. B) IIP3 measurement set-up for 80-GHz tuned amplifier. Before the non-linearity measurements, both W-band power sources need to be powercalibrated either with an external Agilent W-band harm onic mixer 11970W and a spectrum analyzer (SA) E4448A or with a Gilliand W-band power sensor and meter. Thus the available power to the DUT during the measurements is known. Also, the power losses of all the components in the set-up must be me asured separately for de-embedding. As can be seen in Figure 319, the amplifier has an IP1dB of dBm with an input tone at 80 GHz, and an IIP3 of dBm with two input tones at 80 GHz and 80.01GHz.
83 A B Figure 3-18. W-band signal sour ces: A) W-band signal source1. B) W-band signal source2. -68 -58 -48 -38 -28 -18 -8 2 -30-26-22-18-14 Input Power(dBm)Output Power (dBm) 80 GHz Fundamental Tone 79.99 GHz IM Tone -21 dBm IP1dB -11.5 dBm IIP333 Figure 3-19. Measured IIP3 and IP1dB of the LNA under 1.2 V supply and 27 mA drained current. IP1dB is measured with an input at 80 GHz. IIP3 is measured with two inputs of 80 GHz and 80.01 GHz Table 3-1 compares the performance of multi-stage amplifier in the TI 65-nm process with the previously published mm-wav e CMOS tuned amplifiers. Desp ite the use of low leakage transistors with lower fT and fmax, more inexpensive back-end process and higher operating frequency, the performance and power consum ption of the 80-GHz tuned amplifier are
84 comparable to that of the previously reported CMOS mm-wave amplifiers , , and , which all use a cascode topology. Table 3-1. Performance comparisons of CMOS mm-wave low noise amplifiers CMOS Technology Freq. [GHz] Gain [dB] NF [dB] IIP3 [dBm] PDC [mW] Area [mm2] 90/130 GHz fT/fmax, 6M  60 12 8.8 N/A 54(1.5V) 1.3X1.0 120/200 GHz fT/fmax, 9M  58 14.6 4.5(Sim) -6.8 24(1.5V) 0.35X0.4 90/130 GHz fT/fmax 9M  56 24.7 8 -12 72(2.4V) 0.72X0.67 150/110 GHz fT/fmax, 6M (This work) 79 12 9 -11.5 32(1.2V) 0.64X0.38 3.3 Voltage-Controlled Oscillator A voltage-controlled oscillator (VCO) is a funda mental block in RF tr ansceiver. Its output is often used as the LO signal for up/down-conversion mixers and as the clock signal for digital circuits. It can also be used as a frequency modulator. As the name goes, the output frequency of a VCO can be controlled by a volta ge (current) signal. For RF front-ends, the phase noise and output power of the VCO are cr itical characteristics. The SNR of up/down-converted signal is directly influenced by the VCO phase noise and the conversion gain depends on VCO output power over a significant range. There are mainly two types of VCOs: RC-t ype and LC-type. At mm -wave frequencies, currently, an LC-type VCO can oscillate more reliably and provide lower phase noise than an RC-type VCO partially because of the inherent band-pass filtering response of an LC resonant tank. This indicates the importance of the qua lity factor of LC tank to the phase noise. CMOS VCOs have proven their capability in mm-wave operations above 90 GHz . However, the frequency tuning range of these VCOs has been limited to ~2.5%. This is not sufficient to cover the process and temperature variations, and increasi ng the tuning range of
85 mm-wave VCOs is a critical need . With th e CMOS scaling, the unity power gain frequency fmax of MOSFET continues to increase with careful device layout implying that at a certain oscillating frequency, the transist or size required to generate the sufficient negative resistance in a VCO can be reduced. In turn, th e parasitic capacitances associated with the transistor decrease. At the same time, with the scaling of the gate oxide thickness, accumulation-mode MOS varactors may have larger tuning range comparatively with the same gate area. All these point to the possibility of realizing a mm-wave VCO w ith a wider tuning range. However, nano-scale CMOS technologies at 90-nm node and forward support thin and narrow metal interconnects and a moderate number of metal layers for higher dens ity of integration and lower cost. This however limits the Q of inductors and interconnects, and makes impl ementation of mm-wave VCOs challenging. 3.3.1 Design of 94-GHz Voltage-Controlled Oscillator 94 GHz is designated for short distance imaging which has huge potentials in security and medical applications . A 94-GHz fundamental mode VCO is design ed in the TI 65-nm process, targeting to maximize the continuous tuning range of VC O with simple varactors. As shown in Figure 3-20, the VCO is a simple NMOS cross-coupled pair . The NMOS transistor width and length (10.8 m/0.07 m) are sized to obtain sufficient negative resistance and maximum fmax at targeted bias current of ~2-6 mA. A PMOS transistor (240 m/0.5 m) is used as the tail current source. Its relatively la rge gate area reduces the 1/f noise and close-in phase noise of VCO. The PMOS transistor current can be cont rolled in a PLL with the loopfiltered control voltage  to change the DC dr ain voltage of cross-coupled pair so that the VCO tuning range is enhanced -. A 2stage tapered buffer with 3-mA DC bias current/side is used so that the smaller sized 1st stage transistor reduces the capacitive loading of
86 the LC-tank , . The buffe r has its own supply voltage a nd the VCO output power can be adjusted by changing the buffer supply. Figure 3-20. Schematic of the 94 GHz VCO with all transistors sizes in m It is critical to reduce the la yout parasitic capacitances to increase the frequency tuning range. Figure 3-21 shows the layout of a part of the multiple-fingered cross-coupled transistor pair, which shrinks the drain area and increases the distances am ong drain, polysilicon gate and source contacts to reduce Cgs, Cdb, Cds and the coupling capacitances between the two transistor drains. Unlike the layout in , double gate conn ections with double contacts per side are used to lower Rg. This increases the device fmax. Figure 3-21. Part of the cross-coupled tr ansistor pair layout with spacings in m.
87 The LC resonant tank is formed using lumped elements instead of distributed elements to save the circuit area . The tank consists of a single loop ci rcular inductor, a pair of accumulation-mode n-well MOS varactors and the pa rasitic capacitances associated with the NMOS cross-coupled transistor s and buffer transistors. The inductor has ~130 pH inductance and Q of ~10 at 94 GHz in ADS momentum simulations. As discussed in Chapter 2, to satisfy the de nsity requirements, a dense pattern of dummy structures must be added. To prevent the exce ssive loading effects from these to the mm-wave signal paths and inductors while satisfying the metal density rule s, patterned dummy blocks and metal fills must be properly controlled following the rules discussed in Chapter 2. Metal 1-metal 6 dummy metal fills are added in side and outside of the inducto r loop, as shown in Figure 3-22 (A). A B Figure 3-22. LC layouts: A) The loop inductor layout. B) One unit-cell of the varactor layout. The accumulation mode MOS varactor has estimated Q of ~5 at 94 GHz and Cmax to Cmin ratio of ~3 instead of 2 . A unit-cell of th e varactor is shown in Figure 3-22 (B). Each polysilicon gate finger (W/L = 0.64 m/0.18 m) is contacted from two sides with 6 contacts per
88 side to lower the contact resistance (> 60 /contact). The n-well cont act to polysilicon spacing was set to 0.48 m to lower the parasitic capacitance between the two. A micrograph of the VCO is shown in Figure 3-23. The VCO occupies an area of 0.54 x 0.35 mm2 including pads. The oscillator core incl uding the tank and buffe rs is only ~65X65 m2. The dimensions are much less than the ~ 925 m wavelength at 94 GHz in silicon and the lumped models used in simu lations should be acceptable. Figure 3-23. Die photo of the 94-GHz VCO. 3.3.2 Measurements of 94-GHz Voltage-Controlled Oscillator A W-band GSG probe, Agilent 11970W harmonic mixer and E4448A spectrum analyzer are used for the on-chip measurements (single side) (Figure 3-24). The conversion loss of this set-up is measured to be ~50 dB around 94 GHz with Gilliand W-band pow er sensor and power meter. The VCO starts to oscillate at 94 GHz at Vdd = 0.7 V and 1 mA current. For better performance, the VCO was measured at Vdd of 1.5 V and Vbuff of 0.8 V. Figure 3-25 shows the output spectrum at 6 mA VCO core bias current. Figure 3-26 show s the phase noise at the same bias condition is -106 dBc/Hz@10 MHz offset.
89 Figure 3-24. Measurement set-up for the 94-GHz VCO Figure 3-25. Measured VCO out put spectrum before de-embedding (spectrum analyzer RBW 1MHz) with 1.5 V and 6 mA VCO core s upply. After de-embedding the measurement set-up loss, the output power is ~-5 dBm. Figure 3-26. Measured phase noise with 1.5 V and 6 mA VCO core supply with carrier frequency of 94.92 GHz
90 The tuning curves under VCO core bias currents of 2-6 mA are shown in Figure 3-27. At a given bias current, the frequenc y can be tuned over ~3 GHz by varying the varactor tuning voltage (Vtune) between 0 and 1.5 V. Totally, the VCO out put frequency can be varied from 91.8 to 97.4 GHz and the corresponding tuning range is 5.8 %. The output power varies from to dBm (single-sided) within this tuning range. 91 92 93 94 95 96 97 98 0.00.20.40.60.81.01.21.41. 6 Vtune (V)Frequency (GHz) = 7 mA = 10 mA = 12 mAIbiasIbiasIbias Figure 3-27. Voltage-controlled oscillator tuning curves with 1.5 V supply and bias currents of 2, 4, 6 mA for the VCO core For VCOs operating near 90-100 GHz, Table 3-2 shows that despite th e limitations of a moderate number of thin metal layers with stri ct density requirements, the VCO in this work achieves ~ 2X larger tuning ra nge and at least 5 dB higher output power compared to the previously reported VCOs in similar frequenc y range, while almost matching the best phase noise performance. The wider tuning range s hould be able to increa se the tole rance the capacitance and inductance va riations resulting from the proce ss and temperature variations. This is attributed to the careful op timization of transistor sizes and layouts of all devices constituting the VCO tank.
91 Table 3-2. Performance comparisons of CMOS fundamental mode voltage-controlled oscillators around 94-GHz CMOS Technology Freq. [GHz] Phase Noise [dBc/Hz] @ 10 MHz offset PDC [mW] Vdd [V] Tuning [GHz] Output Power [dBm] Area [mm2] 90 nm, fmax ~110 GHz 7M 103.9 -97 120 1.0 N/A -24 N/A 130 nm, fmax 120 GHz 8M  89 -106 15 1.5 2.2 -14 0.5X0.48 130 nm, fmax 120 GHz 8M  98.5/ 105 -102.7/-97.5 15/7 1.5/1.2 2.5/0.2 -18/-25 0.5X0.48 90 nm, fmax 200 GHz 9M  109 -105.2 18 1.5 2.5 -10 0.54X0.36 90 nm, fT/fmax 160/142 GHz 9M  91.5 -107.1 14-87 N/A 2.5 -14 0.62X0.55 65 nm, fT/fmax 150/110 GHz 6M (This work) 94.6 -106 8-14 1.5/0.8 5.6 -8~-4 0.64X0.38 3.4 Down-Conversion Mixer A down-conversion mixer is another fundamental building block in a receiver front-end (specifically, in a heterodyne or homodyne receiver), which transl ates incoming radio frequency (RF) signal to intermediate frequency (IF) ba nd or baseband frequency, and contributes a nonnegligible portion to the overall receiver NF. The latter is especially tr ue in mm-wave receiver systems where the LNA block preceding the mixe r usually has moderate gain (as the 80-GHz tuned amplifier discussed in previous sections) and the mixer itself has relatively high noise figure. At mm-wave frequencies, a lthough a sub-harmonic mixer (SHM) has the advantage of requiring lower local oscillator (LO) frequency generation and transmission scheme, and lower LO-RF feed-through comparing to a fundamental LO mixer - a quadrature LO generation scheme is needed for SHM, which is either difficult for high frequency or consumes
92 relatively large area. Also, a relatively large LO amplitude is required and hence large buffering power is dissipated and LO phase noise could increase from the implic it LO multiplication in SHM. In this section, a fundamental LO down-c onversion mixer that can be fabricated in CMOS is described. 3.4.1 Design of 77-GHz Down-Conversion Mixer A 76-77 GHz CMOS active mixer is designe d in the TI 65-nm process with the fundamental LO and ~5-GHz IF. This is the first CMOS active mixer operating at 76-77 GHz fabricated using low-leakage transistors . Figure 3-28. Schematic of the 77-GHz down-conversion mixer. As shown in Figure 3-28, the mixer consists of a double-balanced Gilb ert-cell core and two on-chip baluns at the RF port a nd LO port to convert external single-ended mm-wave signals to differential signals for testing. The balun is a loop type with 1:1 transfor ming ratio fabricated in metal6, as discussed in Chapter 2 (see Figure 2-11). Custom dummy metal patterns are placed inside and outside of the balun to satisfy the strict metal density rules and mitigate adverse
93 coupling and loading effects from the dummy meta ls, similar to that was done for a loop inductor in the 94-GHz VCO discussed in the last section. Metal capacitors (Ct1-Ct4) are used at the primary and secondary ports of balun to tune with the inductances of the balun and interconnects. For circuit simulations, the Fasthenr y  simulated characteristics of balun are fit to the lumped model in 2-12 (B). Series induc tors are used at the RF and LO ports following the baluns for 50matching. The switching core NMOS sizes are chosen to obtain voltage gain of ~2.5 from the secondary ports of baluns to the gates of NMOS transistors. This lowers the mm-wave LO signal amplitude requirement. A tail inductor of ~200 pH is used to form a resonant tank at the sources of transconductors (M1 and M2), which increases the common m ode rejection ratio (CMRR) of the circuit and mitigates the vol tage headroom reduction due to a tail current source. Series peaking inductors (Lm1 and Lm2) are inserted between the transconductors and switch-quad core to resonate out part of the parasitic capacitan ces of drains of transc onductors and sources of switch-quad transistors, hence ex tending the circuit gain bandwi dth . Proper choices of the peaking inductors boost the mixer conversion ga in by ~2 dB in simulations. The IF output matching circuit consists of tw o spiral inductors and metal cap acitors and is matched to 100 differentially at ~5 GHz. All mm-wave inductors are implemented as line inductors in metal6 like those discussed in Chapter 2. The differential line inductors are separated at least 10 m away to reduces the electro-magnetic coupling, which can change th e inductances. The line le ngths are compensated wherever the line inductances are redu ced due to the coupling effects. The transistor layout is optimized using the same approach discussed in Chapter 2 that reduces parasitic resistances and capacitances for increased fmax. The RF, LO and IF signal paths
94 are carefully drawn to make the layout as symmetric as possible. This is critical to lower the leakage among ports. The differential mixer core topology mitigates the non-ideal ground effect at mm-wave frequencies. The conversion gain (CG), NF and IP1dB of the mixer are simulated assuming an LO of 600 mV p-p at 70 GHz, wher e the maximum CG is ~.5 dB and minimum NF is ~10.5 dB at ~75 GHz. IP1dB is ~ -7 dBm at that frequenc y. A micrograph of the mixer is shown in Figure 3-29. The circuit occupies an area of 0.68 X 0.61 mm2 including bond pads. Figure 3-29. Die photo of the 77-GHz mixer. 3.4.2 Measurements of 77-GHz Down-Conversion Mixer The mixer is measured on chip with a W-band GSG probe, V-band GS probe and 40 GHz GSSG probe with an external hybrid acting as a balun. A Gilliand W-band power meter is used to calibrate the mm-wave signal power into the RF port of mixer. Figure 3-30 shows that the RF and LO port matching below 80 GHz (|S11|) and 76 GHz (|S22|) are less than dB, which are acceptable for measurements. This less than ideal matching is due to the measured capacitance of signal pad with dummy fill structures underneath be ing twice as the estimated value used in the design. The IF port matching is less than dB at 4.5-7 GHz with the exte rnal hybrid connected. The best matching point is ~ 5.7 GHz.
95 Figure 3-30. Measured ports matching of the mi xer with 1.2 V supply and 5 mA bias current. A Gunn oscillator provides a fixed 69.9 GHz LO with ~4 dBm power into the LO port when measuring RF and IF port matching. Figure 3-31 illustrates the set-up for the mixe r conversion gain measurement. Figure 3-32 shows the measured mixer conve rsion gain (CG) unde r varying biasing cu rrents and LO power level after de-embedding LO signal path loss. The LO signal with variable power levels is generated using an Agilent N5260A extension module controlled by an Agilent network analyzer E8361A, which is then cascaded with two V-band power amplifiers to in crease the power level. It is found that the maximum mi xer conversion gain is obtain ed under 1.2 V supply and 5 mA bias current. Figure 3-31. Measurement set-up for the 77-GHz mixer conversion gain.
96 -35 -30 -25 -20 -15 -10 -5 0123456 Mixer Bias Current I (mA) @ V 1.2VCG @ 5.7 GHz IF Output (dB ) LO= 4 dBm LO= 2 dBm LO= 0 dBm LO= -2 dBm LO= -4 dBm LO= -6 dBm LO= -8 dBm LO= -10 dBmddd Figure 3-32. Measured conversi on gain of the mixer under varyi ng bias currents and LO power levels. RF, LO and IF frequencies are fixe d at 75, 69.3 and 5.7 GHz respectively with the LO port loss de-embedded Figure 3-33 plots the measured average mixer conversion gain and SSB NF from repeated measurements with a fixed LO frequency of 69.9 GHz from a Gunn oscillator. The LO power into the mixer LO port is estimated to be ~4 dBm after subtracting the losses of cable and Vband GS probe, and the reflection loss. After de -embbeding the ~6 dB power loss of the LO balun, ~ -2 dBm LO power is delivered to the ga tes of mixer switch-quad tr ansistors. The NF is measured with the 3rd method in , which set-up is s hown in Figure 3-34. Two external PAs/LNAs connected at the mixer output thr ough a hybrid can increase the mixer output noise level above the noise floor of the spectrum anal yzer. This method is suitable for DUT with high gain and/or high NF so that out put noise level of DUT is high relative to the input-referred noise level of the external PAs. In Figure 3-33, the average maximu m conversion gain of dB at 76.3 GHz and the minimum NF of 17.8 dB at 75.5 GHz are observed.
97 -30 -25 -20 -15 -10 -5 737475767778 RF Fre q uenc y ( GHz ) CG (dB)10 15 20 25 30 35NF,NF (dB) CG NF NF 22 Figure 3-33. Measured average mixer conversion gain (CG), SSB NF and NF after deembedding RF input balun loss (NF2) with a fixed 69.9-GHz LO provided by a Gunn oscillator. The LO power into the mixer LO port is ~4 dBm. Figure 3-34. Measurement set-up for the 77-GHz mixer NF. The down-conversion mixer exhibits conversion loss in both simulation and measurements mainly due to the limited device power gain a nd inductor Q at mm-wave frequencies. A more careful simulation of on-chip bal un using HFSS shows two-port loss of ~6.5 dB instead of ~3 dB as originally estimated with Fasthenry due to the mismatch at RF and LO ports. After deembedding the balun loss, the peak conversion gain of the mixer is ~ -1.5 dB and the NF at 75.5 GHz vary from ~10 dB to 13 dB from multiple measurements.
98 The mixer linearity is measured with the setup plotted in Figure 3-35. The mixer has IP1dB of -6.5 dBm at 76.3 GHz and IIP3 of 2.5 dBm (~-13 and ~-4 dBm after de-embedding the balun loss) with two input tones at 76.3 and 76.31 GH z combined with a W-ba nd magic-tee as shown in Figure 3-36. Figure 3-35. Measurement set-up for the 77-GHz mixer IIP3 (and IP1dB by using only one Wband signal source with out the magic-tee). -100 -80 -60 -40 -20 0 -30-20-10010 RF Input Power of 76.3 and 76.31 GHz (dBm)IF Output Power (dBm) 6.4 GHz IF Output 6.39 GHz IF Output IP 6.5 dBm IIP 2.5 dBm 1dB 3 Figure 3-36. Measured IIP3 with two input tones of 76.3 and 76.31 GHz and IP1dB with an input at 76.3 GHz under 1.2 V supply and 5 mA drai n current. The power delivered into the LO port at 69.9 GHz is ~4 dBm. IIP3 of -4 dBm and IP1dB of -13 dBm are obtained after de-embedding the on-chip input balun loss.
99 Leakages (isolations) among ports are also measured from 58 to 78 GHz with the measurement set-up similar to Figure 3-37. V-band GS/SG probe s and cables, an OML V-band harmonic mixer and an Agilent 8563E spectrum an alyzer are used to detect the undesired coupled signal power from one port to another. RF-IF, LO-IF, LO -RF and RF-LO isolations are better than 38, 32, 21 and 31 dB respectively in the measured frequency range, as shown in Figure 3-38. Figure 3-37. Measurement set-up for the 77-GHz mixer LO-RF leakages. -60 -50 -40 -30 -20 -10 556065707580 RF/LO Frequency (GHz)Leakage (dB) RF to IF Leakage LO to IF Leakage RF to LO Leakage LO to RF Leakage Figure 3-38. Measured leakages among mixer ports at VDD=1.2 V and 5 mA drain current. The power delivered into the LO port at 69.9 GHz is ~4 dBm for the RF/LO to IF leakage measurements.
100 Table 3-3 lists the performances of this mixer as well as those of other CMOS mm-wave mixers from the literatures for comparison. Consuming only 6 mW, the highest average conversion power gain of mixer is .5 dB at 76.3 GHz, the aver age NF is 11.3-13.5 dB between 76-77 GHz, and IIP3 and IP1dB are ~-4 and ~-13dBm after de-e mbedding the on-chip RF input balun loss. Such performances are comparable to or better than other published CMOS mm-wave down-conversion mixers at highe r operating frequencies and lo wer LO power. This once again suggests the feasibility of building a complete mm-wave transceiver on a single chip using low leakage transistors in a 65-nm CMOS process for low-cost consumer-oriented applications. Table 3-3. Performance comparisons of st and-alone CMOS mm-wave down-conversion mixer CMOS Technology RF/IF Freq. [GHz] Topology Conversion Gain [dB] NF [dB] IP1dB/ IIP3 [dBm] LO-IF/ LO-RF Isolation [dB] PDC [mW] @ Vdd [V] LO Power [dBm] Area [mm2] 130 nm, 85/135 GHz fT/fmax 8M  60/ 2 Active balanced singlegate -2 11.5 (sim) -3.5/ N/A 10-14/ N/A 2.4 @ 1.2 0 2.72 90 nm, 150/200 GHz fT/fmax 5M  62/ 2 Passive singleended -11.6 11.6 6/16.5 N/A 0 4 4 130 nm, 76/90 GHz fT/fmax 8M  9-50/ 0.01 Active doubly balanced >5 16.4 @15 GHz N/A/ 4.5 @ 30 GHz >20/ >40 97 @ 3.3 5 0.25 65 nm  60/ 2 Passive singly balanced -12.5 12.5 5/ N/A >30 / N/A 0 8.7 0.47 130 nm, 85/90 GHz fT/fmax 8M  35-65/ 3 Active doubly balanced SHM -7.5 17.5 @40 GHz -5 @ 60 GHz/ 9 @ 44 GHz >50/ N/A 91 8 1 65 nm, low leakage NFET >150/ >110 GHz fT/fmax 6M (This work) 76/ 6 Active doubly balanced -8 (-1.5 w/o balun loss) 17.8 (11.3 w/o balun loss) -6.5 /2.5 (~-13/4 w/o balun loss) >32/ >21 6 @ 1.2 ~4 on the LO port, ~ -2 on the mixer switchgates w/o balun loss 0.42
101 3.5 Summary In this chapter, several fundamental blocks in mm-wave receiver front-end implemented in the TI 65-nm bulk CMOS process are presented. First, the design flow of an 80-GHz tuned amplifier with six stages is given, followed by the discussion of nonideal effects caused by onchip ground and supply network, and silicon s ubstrate coupling. Meas urements show the amplifier achieves 12 dB gain at 79 GHz and NF less than 10.5 dB within the 6 GHz 3-dB bandwidth. IP1dB of -21 dBm and IIP3 of -11 dBm are measured. Next, a compact 94-GHz LCtype VCO with 5.8% tuning range is presented. Its output power reaches -5 dBm and the phase noise is -106 dBc/Hz @10 MH z offset. Finally, a 76-77 GHz down-conversion mixer with fundamental LO and 5 GHz IF is designed and implemented. Measurements show the highest average conversion power gain of mixe r is .5 dB at 76.3 GHz, the averagenoise figure is 11.3-13.5 dB between 76-77 GHz, and IIP3 and IP1dB are ~-4 and ~-13dBm after de-embedding the on-chip RF input balun loss. The simulated and measured results of mm-wave circuit blocks indicate the feasibility of implementing an entire mm-wav e receiver using low-leakage tr ansistors of bulk CMOS, which implies low-cost and high integration.
102 CHAPTER 4 APPLICATIONS FOR CMOS MILLIMETER-W AVE RECEIVER FRONT-END CIRCUITS 4.1 Introduction The feasibility study for mm-wav e receiver circuit blocks in Chapter 3 shows that it is possible to build a complete mm-wave trans ceiver in the 65-nm low leakage bulk CMOS process. Such transceiver can be used for many applications at mm-wave frequencies. Here, two applications are considered. The first one is a mm-wave wireless interconnect system targeting to alleviate some issues of conventional wire-b ased interconnect system, namely, complex I/O wiring, larger package area and limited data band width. This design is at system-level based on circuit block performances in Chapter 3. The se cond application is a 77 -GHz automobile radar system. Receiver circuit design and experiments are the focus of that. In the following sections, this application will be discussed. However, the system design details for mm-wave wireless interconnects are given in Appendix A. 4.2 Introduction to 77-GHz Automobile Radar Radar stands for radio detection and ranging. Radar systems are very useful in a multitude of applications such as air defense systems, na vigation, weather monitoring, and object detection . Simply put, all of these radar systems function on the same principle of radiating electromagnetic energy away from an RF sour ce and monitoring any reflected signals from objects. The angle at which a receive antenna is pointed, the time delay between transmitted and received signals, and any frequency variation between the transmitted and received signals, together are used to determine the location, m oving direction and speed of the detected object. The received signal can also be an alyzed to obtain the properties of the object such as the shape, size, and constitutes. In recent years, radar-ba sed driver assistance sy stems such as adaptive cruise control (ACC) and collision avoidance have been introduced into the automobile market
103 to enhance the safety and comfort of driving. Most of these systems rely on microwave or mmwave radar to obtain the necessary information about the driving e nvironment . . It is estimated more than 10 radars per vehicle are need ed for various needs, as shown in Figure 4-1. Figure 4-1. Vehicle with multiple radar modules for various functi ons to make driving more safe and comfortable Considering the sales of millions of units per year and the high probability of lowering life and property losses on the road, it is not surprising that the autom obile radar applications attract
104 strong interests from both the car and electronics manufacturers. Compared to the radar system at lower frequency bands such as 5 GHz and 24 GHz, mm-wave system will need a smaller antenna to obtain the needed directivity for higher a ngular resolution, thus making the radar more compact and easier for installation in limited spac e available in a car, such as the space behind the front bumper. Cost could also be lowered. On the other ha nd, higher frequencies in the EM wave spectrum, such as infrared and optical wavelengths, would allow even narrower beam width than at mm-wave frequencies for the same size of antenna. However, these shorter wavelengths waves suffer from much larger attenuation during propa gation especially from fog, dust, smoke and others. 76-77 GHz in the mm-wav e region is one frequency band designated for automobile radar applications, which has relative ly low excess propagation loss, as can be seen in 1-2 of Chapter 1. More recently, a shortrange wide band radar system (SRR) using 77-81 GHz is also proposed . For simplicity, we w ill refer the automobile radar applications in both bands as 77-GHz radar in the following sections. Up to now, automobile radar circuits are mo stly fabricated in III -V technology, which contributes to the high cost and impedes the wide use. With the speed improvement of CMOS technology, low-cost digital CMOS devices have been used to dem onstrate circuits operating in mm-wave band as discussed in Chapters 2 and 3. These suggest the po ssibility of highly integrated automobile radar chips with much lower cost than those based on III-V technology, even at frequency as high as 80 GHz. To further study this opportunity, a 77-GHz pulsed radar transcei ver architecture is proposed in this chapter and th e corresponding receiver front-end circ uits are designed in the TI 65-nm CMOS technology. First, an overview for the proposed rada r system is presented in the next section.
105 4.3 Overview of the 77-GHz Radar System For the proposed 77-GHz CMOS automobile radar system, it is assumed that a multi-beam high gain antenna from Roadeye Ltd.  is used for both transmission and reception. The front and rear view of antenna a nd its radiation pattern are shown in Figure 4-2. The three narrow main beams in the center are for long range cove rage and the two wider side beams are for wide angle coverage. Transmitter and receiver circui ts of the radar can be mounted behind the Roadeye antenna, forming a compact radar module. Figure 4-2. Multi-beam antenna and its radia tion pattern for the proposed 77-GHz CMOS radar system. To lower cost and to exploit the benefits of CMOS integration, transmitter and receiver should be built on one die. Since 76-77 GHz is the operating range of the radar system, transmitting and receiving in the only 1 GHz band at such high frequency and close distance between a transmitter (TX) and a receiver (RX) will require impossibly high isolation among TX and RX in a continuous wave rada r as introduced in [ 109] to avoid serious mutual interference, resulting from TX output signals being so much higher than the desired reflected signal the RX needs to detect. Hence a pulsed radar architecture is chosen for the target ed radar system, which transmits a 77-GHz wave modulated by a short sq uare pulse, then turns off the transmitter and
106 turns on the receiver to listen for possible reflected signals. The time division transmitting and receiving eliminates the mutual interference pr oblem between TX and RX. The way pulsed radar works is similar to a time-domain reflectometer many RF designers are familiar with. In pulsed radar, the spatial resolution is de termined by the time delay between the starting moments of transmitter radiation and receiver sensing. Thus it is ultimately bounded by the transmitted pulse width. To reduce the potential mutual interferences among multiple working pulsed radars in a real driving environment, pseudo-random coded sequences of pulses can be used to distinguish different vehicles. Based on these, the proposed radar system speci fication is tabulated in Table 4-1. As seen in Table 4-1, the pulsed radar ta rgets a detection range of 1-150 m with pulse widths between 2.5 and 25 ns depending on the object distance from the radar, which translates to an RF signal bandwidth of 800 and 80 MHz for a pulse-transmitti ng rate of 10 Hz. These narrow pulses allow resolutions of 1 m for the range of 1-35 m and 3.75 m for 35-150 m. Also, pulse averaging and a digital signal processing t echniques can be used in the syst em to increase the detected signal SNR (corresponding to detection e rror probability). The maximum number of integrated pulses depends on the object distance. Table 4-1. System specification for th e proposed 77-GHz pulsed radar system Features Specifications Range (m) 1-150 Resolution (m) 1-3.75 Pulse Width (ns) 2.5 25 Pulse Signal Bandwidth (MHz) 80 -800 Pout (dBm) 0 10 Averaging Pulse Number 10-5,000 Refresh Rate (Hz) 10 Receive Time Slot ( S) 1.5
107 Assuming TX output power of 0-10 dBm, RX NF of 10 dB, minimum output SNR of 13 dB (corresponding to 10-8 detection error probability), link ma rgin of 5 dB and detected object radar cross section area of 1 m2, the radar equation  is used to calculate the detection range Equation 4-1 for far and short range scenarios resp ectively. The results are summarized in Table 4-2 and 4-3. sys tLSNRkTBF nGP Rmin 3 22 max)4( (4-1) Table 4-2. Pulsed radar system range calculation for long distance detection Items Values Items Values TX output power Pt 10 dBm B (bandwidth) 79 dB-Hz Antenna gain (G) 25 dB (receive) + 30 dB (transmit) Pulse width 25 nS Radar cross section(s) 0 dB F (noise factor) 10 dB @ 77-GHz -82.1 dB SNRmin 13 dB n (number of integrated pulses) 37 dB Link margin (Lsys) 5 dB KT -174 dBm/Hz Max. range Rmax 157 m Resolution 3.75 m RX received power Pr @ Rmax -104 dBm Table 4-3. Pulsed radar system range calculation for short distance detection Items Values Items Values TX output power Pt 0 dBm B (bandwidth) 89 dB-Hz Antenna gain (G) 25 dB (receive) + 30 dB (transmit) Pulse width 2.5 nS Radar cross section(s) 0 dB F (noise factor) 10 dB @ 77-GHz -82.1 dB SNRmin 13 dB n (number of integrated pulses) 10 dB Link Margin (Lsys) 5 dB KT -174 dBm/Hz Max. range Rmax 11 m Resolution 1 m RX received power Pr @ Rmax -69 dBm
108 Notice that for narrower pulses (higher spatial resolution), the signal bandwidth is larger, implying the background thermal noise power at the RX is higher and the maximum range will be reduced. 4.4 Design and Measurement of an 80-GHz Down-Converter Before a further study for the 77-GHz pulsed ra dar transceiver archit ecture and receiver circuits, an 80-GHz down-converter aiming to demonstrate the f easibility of building a 77-GHz receiver is designed and implemen ted using the low leakage transi stors of TI 65-nm bulk CMOS process. The design of the 80-GHz down-c onverter is as follows: A highe r peak gain and narrower 3-dB bandwidth design of the 80GHz tuned amplifier with its performance shown in Section 3.2 (Figure 3-10 (A)) is cascaded with a 77-GHz pseudo-differentia l down-conversion mixer  similar to the double-balanced mixer described in Section 3.4. Figure 43 illustrates that the interface between the circuit blocks is changed from 50matching to a series LC matching network where L is provided by the parasitic inductance of the relatively long interconnect. The balun at the RF port of 77-GHz mi xer in Chapter 3 is removed because of its power loss (~ 3 dB) at mm-wave frequencies. Hence, the mixer is mo dified from the differen tial input-differential output topology to a single-end ed input-differential output topology (pseudo-differential) as shown in Figure 4-4. Tables 4-4 and 4-5 list the component values in Fi gure 4-3 and Figure 4-4, respectively.
109 Figure 4-3. Schematic of the tuned amplifier in the 80-GHz down-converter. Figure 4-4. Schematic of the pseudo diffe rential mixer in the 80-GHz down-converter.
110 Table 4-4. Component values of the tuned amplifier in the 80-GHz down-converter M1-M6 (W) ( m) Pad Cap. (fF) Cbypass (pF) C1 (fF) 18 80 1 190 Lg1 (pH) Ls1 (pH) Ld1 (pH) C2 (fF) 190 20 100 130 Lg2 (pH) Ls2 (pH) Ld2 (pH) C3 (fF) 110 15 90 130 Lg3 (pH) Ls3 (pH) Ld3 (pH) C4 (fF) 110 10 100 130 Lg4 (pH) Ls4 (pH) Ld4 (pH) C5 (fF) 110 10 80 130 Lg5 (pH) Ls5 (pH) Ld5 (pH) C6 (fF) 110 5 80 120 Lg6 (pH) Ls6 (pH) Ld6 (pH) 110 5 110 Table 4-5. Component values of the pseudo differential mixer in the 80-GHz down-converter M7-M12 (W) ( m) Lg_mix (pH) Cg_mix (pF) Ls_mix1,2 (pH) 32 80 1 10 Ltail (pH) Lm1,2 (pH) Lp1,2 (pH) Ct1 (fF) 200 100 120 80 Ct2 (fF) Cdif_mix1,2 (fF) Ldif_mix1,2 (pH) Cbypass_mix(pF) 140 200 360 24 Transistors M7-M12 are 32 m wide. The highest mixer conv ersion gain is achieved with 600 mV p-p LO swing on the gates of M9-M12. An 1:1 on-chip balun built with two overlapping single loops in metal6 and pad aluminum layer is used at the LO port of the mixer to convert external single-ended LO signals to differential LO signals, sim ilar to that in . The balun has ~3 dB loss in HFSS simulati ons at 60 GHz. Metal capacitors (Ct1-Ct4) are used to tune the balun to ~60 GHz. Series inductors Lp1 and Lp2 boost the LO swing at the gates of NMOS
111 switching transistors (M9-M12), compensating part of the LO power loss from the balun and relaxing the external V-band LO signal drive requirement. M7 and M8 work as an active balun. The tail inductor, Ltail forms a resonant tank with the source degeneration inductors (LS_mix1 and LS_mix2) and parasitic capacitances of M7 and M8. The tail inductor increases the comm on mode rejection of mixer. The inductors make the input impedance seen at the gate of M7 higher and its gate to source voltage gain closer to the theoretical value of for increased single-ende d to differential conversion gain. The simulated amplitude imbalance between th e two IF outputs is ~1.5 dB and can be compensated in differential circuits following the mixer. Lm1 and Lm2 are inserted between M7, M8 and the switch-quad to boost the mixer conversion gain by ~2 dB . In the mixer layout, the symmetry of signal paths is critical to lower the leakage among ports. In simulations, the mixer shows a peak convers ion gain of .5 dB at 79.5 GHz and 11-dB minimum NF. For the whole down-converter, the simulated conversion gain and NF at VDD=1.2V are 19 dB and 7.7 dB at 80 GHz w ith 600-mV p-p LO on the gates of NMOS switching transistors. A die photo of the down convert er is shown in Figure 4-5. The down-converter area is 1.07X0.55 mm2 including bond pads. Figures 4-5 (A), (B ) and (C) are the photographs of downconverter, and zoom-ins for the tuned amplifier and mixer, respectively. As can be seen, line inductors are used on the drains and sources of tr ansistors in the tuned amplifier because of their higher Q factor, while spiral inductors with cust om dummy metal fills are used in other places for more compact layout. The on-chip metal1-6 stacked ground plane and dense dummy metal fills added by the foundry can also be seen in Figure 4-5.
112 A B Figure 4-5. The 80-GHz down-converter die photo A) The whole down-c onverter. B) Close-up of the tuned am plifier circuit. C) Close-up of the mixer circuit.
113 C Figure 4-5. Continued With measurement set-ups similar to that ha ve been described in Chapter 3 for mixer measurements, the down-converter is characterized on chip wi th a W-band GSG probe, a V-band GS probe and a 40-GHz GSSG probe with an ex ternal hybrid connected at the IF port. The matching characteristics at RF, LO and IF ports are measured with an Agilent PNA E8361A and two N5260A frequency extension modules. The measurements a nd simulations of frequency tuning for RF, LO and IF port matching are close, but the actual matching is not as good as the simulated, as shown in Figure 4-6. This is likely due to underestimation of signal pad losses through EM fields coupling to dummy fill stru ctures underneath and around the pads, which lowers the Q of pads and hence the matching netw orks, since pads are an integral part of the matching networks. Such effects are veri fied with simulations. The measured |S11|, |S22| and |S33| around 80, 60 and 20 GHz respectively are less than dB.
114 -16 -12 -8 -4 0 4 020406080100 Frequency (GHz)|S11|, |S22| and |S33| (dB) |S | @RF Port |S | @LO Port |S | @IF Port Sim. |S | @RF Port Sim. |S | @LO Port Sim. |S | @IF Port11 22 33 11 22 33 Figure 4-6. Measured and simulated impe dance m a tching of the down-converter. The conversion gain of down-converter is ch aracterized with an Agilent E4448A spectrum analyzer at VDD=1.2 V and 1.5 V. Figure 4-7 shows the measured conversion gain and SSB NF of the down-converter with 60-GHz LO signa l generated by an N5260A extension module cascaded with two Terabeam V-band power ampl ifiers. The power delivered to the LO port measured with an OML V-band harmonic mixer a nd a spectrum analyzer is ~4 dBm. It is estimated that ~1 dBm LO power is available at the output of on-ch ip balun. As can be seen, the down-converter achieves maximum conversion gain of 16 dB and minimum NF of 9 dB at 80 GHz and VDD=1.2 V. The 3-dB bandwidth is ~5 GHz around 80 GHz. At 1.5 V and 63-mW power consumption, the maximum conversion gain of 18 dB and minimum NF of 8.4 dB are achieved. The measured peak conversion gain an d lowest NF at 1.2 V supply and 53 mW power consumption are 3 dB lower and 1.3 dB higher, re spectively than the simulated results including the non-ideal ground effects. The difference may be attributed to the component model inaccuracy, e.g., the transistor model parame ters being used are intended for low GHz applications. Figure 4-8 plots the conversion gain at varyi ng LO power levels. The peak
115 conversion gain increases from ~ 12 dB to ~16 dB when the power available to the LO port of down-converter is increased from -2 dBm to 4 dBm (-5 dBm to 1 dBm after de-embedding the on-chip balun loss of ~ 3 dB). 0 5 10 15 20 25 30 757779818385 RF Frequency (GHz)Conv. Gain (dB)7 9.5 12 14.5 17 19.5 22NF (dB) Conv.Gain @1.2V Conv.Gain @1.5V Sim.Conv.Gain @1.2V NF @1.2 V NF @1.5 V Sim.NF @1.2V Figure 4-7. Measured conversion gain and NF of the down-converter for fixed LO frequency of 60 GHz. -4 0 4 8 12 16 757779818385 RF Frequency (GHz)Conv. Gain (dB) Conv.Gain@ 4 dBm LO Conv.Gain@ 1 dBm LO Conv.Gain@ -2 dBm LO Conv.Gain@ -5 dBm LO Figure 4-8. Measured conversion gain of the down -converter at varying pow er levels of the 60GHz LO del i vered to the LO port.
116 For IIP3 measurement, two test tones at 81 and 81.01 GHz are gernerated using two N5260A extension modules and combined with a DualComm W-band ma gic-tee. A DualComm Gunn oscillator delivers ~4 dBm LO power at 61 GHz to the LO port of the circuit after deembedding the loss of V-band cables, probe and circuit return loss. Figure 4-9 shows the measured IIP3 and IP1dB of the down-converter are -16 dBm and -26 dBm at VDD=1.2V. LO-RF (input of tuned amplifier) and LO-IF leakages ar e measured to be less than -61 and -24 dB, respectively (Figure 4-10). -95 -75 -55 -35 -15 5 -50-40-30-20-10 RF Input Power of 81 and 81.01 GHz (dBm)IF Output Power (dBm) 20.01 GHz IF Output 20.02 GHz IF Output IP1dB -26 dBm IIP3 -16 dBm Figure 4-9. Measured IIP3 and IP1dB of the down-converter at 81 GHz. -95 -80 -65 -50 -35 -20 45505560657075 LO Frequency (GHz)Leakage (dB) LO-->IF LO-->RF Figure 4-10. Measured LO-RF and LO-IF leakages of the down-converter.
117 Table 4-6. Comparisons of W-band CMOS receivers W-Band Receivers    This Work CMOS technology 90-nm, fmax>120 GHz 65-nm, low Vth fmax/fT 240/165 GHz @ Vds 0.7 V, 7M 65-nm, low Vth, fmax/fT 280/195 GHz @ Vds 1.2 V, 7M 65-nm, high Vth (>0.6 V), fmax/fT >110/>150 GHz, 6M (Dummy metal fills enforced on all passives) RF/IF freq. [GHz] 57-77/0-2 75-91/0-10 76-95/0-18 77-82/17-22 Integration RX (LNA+RF Mix. + IF Mix. + IF Amp.) + PLL RX (LNA+ Mix. + IF Amp.) RX (LNA+Mix. +IF Amp) +VCO +Div2 RX (LNA+Mix.) Conversion gain [dB] 19 @ 77 GHz (voltage gain) 13 @ 80 GHz 7.5 (1.2 V)/ 2 (1.5 V) @ 85 GHz 16 (1.2 V)/18 (1.5 V) @ 80 GHz NF [dB] N/A 8.5-10 (DSB) 7-10 (1.2 V)/6-9 (1.5 V) DSB) 9-11(1.2V)/8.4-10.5 (1.5V) (SSB) IP1dB/ IIP3 [dBm] -16/ N/A -16/ N/A -18/ N/A -26/-16 LO-IF/LO-RF leakage [dB] N/A N/A /< -60 N/A <-23/<-60 LO power [dBm] N/A 5 N/A 4 (1 dBm after deembedding on-chip balun loss) PDC [mW] @ Vdd 60(RX)+72(PL L) @1.5 V 89 @1.5 V 78(RX) @1.5 V + 128 (VCO+Div2) @ 1.2 V 53 @1.2 V 63 @1.5 V Area [mm2] 1.0x1.0 0.46x0.5 1.09x0.6 1.07x0.55 Table 4-6 summarizes the performance of this down-converter as well as those of other published CMOS W-band receiver front-end for comp arison. It shows that at lower DC supply voltage, the down-converter in this pape r fabricated using low-leakage (high Vth) transistors of a lower cost 65-nm CMOS process with a fewer number of metal layers achieves a higher conversion gain and better SSB NF than the two RXs fabricated using low Vth general purpose
118 transistors in 65-nm CMOS , . These ar e attributed to the use of a 6-stage tuned amplifier, elimination of lossy balun in the RF signal path, optimization of both active and passive components, especially the latter in the presence of dummy metal fills, and optimization of the bandwidth for the radar application. Due to the high Vth of the transistors and relatively high gain of the tuned amplifier, the linearity of the down-converter is lower than the RXs in , . Overall, the down-c onverter achieves comparable perf ormance as those fabricated using the general purpose (low Vth) NMOS transistors of a 65-nm CMOS technology. With the successful demonstration of the 80-GHz down-converter, we will continue to investigate the feasibility of a more complete integration of mm-wave receiver front-end for 77GHz band radar application in the low-cost 65-nm CMOS technology. 4.5 Radio Architecture for the 77-GHz Pulsed Radar System For the proposed 77-GHz pulsed radar system, hi gh performance/cost ratio is the key to the success in the high volume market Integrating as many necessary circuit blocks in a single CMOS chip should lower external component counts and packaging cost. With this goal in mind, two basic architectures of RF transceivers, dire ct conversion and dual conversion are evaluated. /2 /128 51 GHz 51 GHz 25.5 GHz 76.5 GHz On-chip Dipole Antenna 25.5 GHz 76.5 GHz PA PFD LF RX BB Output TX BB Input Modulating PA fREF 76.5 GHz Offset Mixer 76.5 GHz LNA 25.5 GHz Tuned Amp 77-GHz Radar RF SOC in CMOS AC Coupling Cap /2 /128 51 GHz 51 GHz 25.5 GHz 76.5 GHz On-chip Dipole Antenna 25.5 GHz 76.5 GHz PA PFD LF RX BB Output TX BB Input Modulating PA fREF 76.5 GHz Offset Mixer 76.5 GHz LNA 25.5 GHz Tuned Amp 77-GHz Radar RF SOC in CMOS AC Coupling Cap Figure 4-11. Simplified block diagram of the CMOS heterodyne radar transceiver.
119 A dual conversion heterodyne arch itecture utilizing a single PLL (Figure 4-11) is proposed to alleviate several issu es with the direct conversion architecture. The proposed architecture can effectively eliminate the problem s of VCO being pulled by the pul se modulated 76.5-GHz carrier from the transmitter and the LO leakage to the receiver which can swamp out the wanted received signal, as well as relaxing the maximum operating fre quency requirement of PLL to 51 GHz, which can not only improve the tuning ra nge and phase noise performance of the synthesizer but also allow the use of compact and wide-locking range static dividers. For the dual conversion architecture, potentially, there are mo re flexibility for the trade-off of individual circuit block performance in the RX chain to meet the overall system requirements. This is an important advantage due to th e limited performances of CM OS mm-wave circuit blocks. The 51-GHz PLL output can be divided by two to generate the 25.5-GHz signal as the LO for the second frequency down-c onversion in the RX. Also, the 51 and 25.5 GHz signal can be offset-mixed to generate the 76.5 GHz carrier needed in the TX. The lower sideband resulting from the offset mixing is 51 GHz below the want ed carrier and can be easily filtered out. 25.5 GHz LO 76.5 GHz RX Input RX BB Output 76.5 GHz LNA 25.5 GHz Tuned Amp 51GHz LO AC Coupling Cap 76.5 GHz Mixer 25.5 GHz Mixer 1 2 325.5 GHz LO Buffer 51GHz LO Buffer 25.5 GHz LO 76.5 GHz RX Input RX BB Output 76.5 GHz LNA 25.5 GHz Tuned Amp 51GHz LO AC Coupling Cap 76.5 GHz Mixer 25.5 GHz Mixer 1 2 325.5 GHz LO Buffer 51GHz LO Buffer Figure 4-12. RX block diagram showi ng the main 25.5-GHz LO leakage paths In the frequency plan, the received 76.5GHz signal has image fr equency at ~25.5-GHz. Due to the high IF of 25.5-GHz for this dual conversion architecture, the received image interference can be effectively attenuated by the bandpass filtering effect of the mm-wave LNA
120 and mixer. However, at the input of the 25.5GHz down-conversion mixer, there is not only the desired down-converted 25.5-GHz signal but also leakages of the 25.5-GHz LO. The 25.5-GHz LO can take several main paths to end up at the input of the 25.5-GHz mixer (Figure 4-12): (1) Leak to the 25.5-GHz mixer RF inpu t due to its limited LO-RF isolat ion. (2) Leak to the input of the 25.5-GHz tuned amplifier. (3) Leak to the in put of the 76.5 GHz mixer, and mix with the 51GHz LO to generate the 25.5 GHz signal. Thes e leakages may be from the EM near-field coupling and EM wave radiation th rough the signal paths of circu its, conductive silicon substrate and non-ideal on-chip power and ground lines. Fo r several of these cas es, the problem is compounded by the subsequent amplification of the IF amplifier. This is a serious concern that must be addressed during the de sign of the receiver circuit. Notice that in our actual circuit design in the following three sections, all the circuit blocks are intentionally designed to opera te at ~10% higher frequencies than those shown in Figure 4-11 to provide an adequate margin for frequency tu ning down due to variations and underestimated parasitic capacitances. FRONT VIEW TOP VIEW FRONT VIEW ANTENNA LENGTH 50 LUMPED PORT AIR BOX RECTANGULAR WAVEGUIDE TO HIGH GAIN ANTENNA METAL PLANE BENEATH THE SUBSTRATE PCB SUBSTRATE SILICON SUBSTRATE ANT WIDTH 50 WAVE PORT BIRDS EYE VIEW FRONT VIEW TOP VIEW FRONT VIEW ANTENNA LENGTH 50 LUMPED PORT AIR BOX RECTANGULAR WAVEGUIDE TO HIGH GAIN ANTENNA METAL PLANE BENEATH THE SUBSTRATE PCB SUBSTRATE SILICON SUBSTRATE ANT WIDTH 50 WAVE PORT BIRDS EYE VIEW Figure 4-13. Three-dimension modeling of the tr ansition from the 76.5-GHz on-chip dipole feed to the waveguide section of a high gain antenna
121 As discussed, it is expected that a T/R switch in the TI 65-nm process at 77-GHz will have an insertion loss 4-5 dB based on the measured insertion loss of ~4.5 dB at 67 GHz. This insertion loss is fundamentally limited by the relatively large losses in passive components such as inductors and capacitors, and PN junctions of drain and source of transistors in the low-cost digital process. This degrades the TX output power and RX sensitivity, limiting the link margin of the system. To avoid the sw itch loss, in the ra dio architecture propos ed in Figure 4-11, separate on-chip feeds (dipoles) for waveguide s at 76.5 GHz are integrated for TX/RX. As shown in Figure 4-13, the TX feed transmits sign al to and the RX feed receives signal from a waveguide of the high gain Roadeye antenna  in Figure 4-2. For this approach, there is no need for mm-wave off-chip connections through a package and a substr ate. This eases the assembly of the radar module and lowers the cost Studies of this on-ch ip feed (dipole) to waveguide transition have been conducted and EM simulations show ~3.5 dB loss for the dipole feed-to-waveguide transition around 77-GHz. This loss is expected to be ~0.5 dB different from that of a system using the best mm-wave T/R switch and off-chip waveguide feeds connected through bond wires-transm ission lines . 4.6 Design of Receiver Chain for the 77-GHz Pulsed Radar Transceiver The 77-GHz radar receiver chain is designe d in the TI 65-nm CMOS process, which includes a 82.5-GHz RF tuned amplifier and an active down-conversion mixer (Figure 4-14) followed by a 27.5-GHz IF tuned amplifier and a passive down-conversion mixer, and finally baseband HPF and current buffer (Figure 4-15). A differential (pseudo-differential) topology is desirable for its reje ction of common mode noise and non-ideal ground/supply effects as discussed in Chapter 3. This feature is especially important in the noisy on-chip environment where circuit blocks can interfere with each other through undesired coupling. Besides, the on-chip dipole feed can be conveniently connected with
122 the first differential amplifier stage of the rece iver front-end and there is no need for singleended to differential conversion. However, since only single-ende d (SE) probes are available at 77-GHz, for the first RF amplifier, it is designed as three SE CS amplifier stages cascading with Figure 4-14. Receiver mm-wave front-end schema tic (RF stages) of the 77-GHz radar system three differential CS amplifier stages with the SE to differential conversion accomplished by a 77GHz line balun discussed in Chapter 2. The pr imary winding of the line balun serves as an inductive load for the third stag e SE CS amplifier. The first thr ee stages of SE CS amplifiers facilitate SE GSG probe testing and suppress the RX NF degradation from the line balun loss. The SE CS amplifiers can be changed to differe ntial CS amplifiers and the line balun can be removed if the receiver chain is inte grated with an on-chip dipole feed. All the CS amplifier design follows the me thod discussed in Chapter 3. The commonmode rejection ratio (CMRR) of the 3 differentia l CS amplifier stages ar e enhanced by inserting inductors between the co mmon source of CS amplifier and on -chip ground. In simulations, the RF amplifier achieves a differential voltage ga in of ~14 dB and good input match is obtained between 77 and 85 GHz. The CMRR is simulated to be ~40 dB and reverse gain is ~ -50 dB. The
123 amplitude imbalance between the two differential output is less than 0.5 dB and the phase error is less than 10 degree. Figure 4-15. Receiver IF and BB stages schematic of the 77-GHz radar system The RF down-conversion mixer is a double-balanced Gilbert ce ll and its design is similar to the 77-GHz down-conversion mixer core design in Chapter 3. The LO frequency for the mixer is ~55 GHz. There is no need for an on-chip ba lun at the RF input of the mixer since the amplifier preceding it already provides differential signals, but the LO input of the mixer still requires an on-chip balun if the receiver is test ed with an external LO Notice that the mixer switch quad transistor sizes should not be excessively la rge so that he drive requirement of mmwave LO can be lowered. In simulations, this mi xer obtains a voltage conv ersion gain of ~ -2.5 dB, LO-RF isolation of ~20 dB and CMRR of ~2 0 dB. The RF amplifier and mixer combination shows NF of ~ 8.7 dB in simulations. The IF amplifier is a cascade of 4 stages of differential cascode amplifiers, which should increase the reverse isolation at 26 GHz. The tran sistor sizes are selected for high gain and low NF. The whole IF amplifier can provide ~20 dB vo ltage gain and its CMRR is more than 60 dB.
124 The block following the IF amplifier is an IF mixer, which down-converts the incoming signal centering at ~27 GHz to baseband. A doublebalanced passive mixer topology is chosen for the IF mixer. By trading off conversion gain, the relatively high linearity and low flicker noise of the passive mixer compared to active mi xer help maintain the output signal quality at baseband. In simulations, the 26 GHz passive mixe r achieves voltage conver sion gain of ~-6 dB The baseband differential signal output from th e passive mixer will then enter a simple HPF consisting of a series metal coupling capacitor and a resistor shunt to ground, and then to a current source follower that acts as a signal buffer to an external load. The HPF 3-dB bandwidth is set to ~1 MHz to filter out any other DC offs et and low frequency noise present. The desired received signal will only lose a very small portion of its energy through the HPF because the signal is supposed to occupy at least 40 MHz baseband bandwidt h (for pulse width of 25-ns) according to the pulsed radar system design in Section 4.3. The coupling cap and shunt resistor values are co-optimized with the current source follower to obtain the best si gnal gain from the passive mixer to the off-chip 50load emulating the input resistance of test equipment. The transistor sizes of current source follower are selected for a good compromise between larg e driving current and small loading to the passive mixer. In simulations, the combina tion of baseband HPF and current source buffer driving 50 ohm load shows ~ -10 dB voltage ga in and bandpass characteristics between ~1MHz and 3 GHz. For the entire receiver chain, the conversion gain is ~15 dB and the NF is ~ 10.5 dB in simulations, assuming the 55 and 27.5 GHz LO signal swings are both 60 0 mV peak-to-peak. With the preliminary RX chain circuit designs and simulations in place, we are ready to look back at the 27.5 GHz LO leakage problem disc ussed in the previous section. The amplitude
125 Table 4-7. Estimation of 27.5GHz local oscillation leakage at the input of 27.5-GHz downconversion mixer Simulated circuit block Forw ard gain/ conversion gain (dB) Reverse loss/LO-RF isolation (dB) 82.5-GHz LNA -50 @27.5 GHz 100 @27.5 GHz 82.5-GHz down-conversion mixer (Mixer 1) -26 @27.5 GHz (CMRR 20 dB) 100 @27.5 GHz 27.5-GHz tuned amplifier (Tuned Amp) 21 @27.5 GHz (CMRR 50 dB) 40 @ 27.5 GHz 27.5-GHz down-conversion mixer (Mixer 2) -6 @ 27.5 GHz (IV1dB ~0.8 V) 30 @27.5 GHz Estimated coupling loss (dB) 24 (from 27.5 GHz LO buffer output to the 27.5-GHz tuned amplifier input) 30 (from 27.5 GHz LO buffer output to the 76.5-GHz mixer input) Simulated 27.5-GHz LO buffer output amplitude 0 dBV (1 V) 27.5-GHz LO leakage of Path 1LO buffer output Mixer 2 LO port Mixer 2 input 0-30 = -30 dBV 27.5-GHz LO leakage of Path 2LO buffer outputTuned Amp inputMixer 2 input 0-24-20+21 = -23 dBV Assume 1/10th of the coupled voltage are differential signal 27.5-GHz LO leakage of Path 3: LO buffer output Mixer 1 inputTuned Amp inputMixer 2 input 0-30-6-26+21= -47 dBV Assume 1/10th of the coupled voltage are differential signal 27.5-GHz LO total leakage of Path 1,2 & 3 Amplitude 0.07 V < 0.8 V (IV1dB of Mixer 2) of total 27.5-GHz LO leakage at the input of 27.5-GHz mixer can be roughly estimated to be ~ 23 dBV (0.07 V) based on the simulated circui t performances and estimated coupling losses from the past measurements on circuits fabricat ed in the 65-nm proce ss (Table 4-7). The 27.5-
126 GHz mixer has a simulated IV1dB of ~0.8 V and the leaked 27.5 -GHz signal at the input of the mixer will be down-converted to become DC offset/low frequency noise with a bandwidth of ~1MHz at the output of the mixer, which is easily removed by the RC HPF at the baseband. So this leakage should not cause nonlinear dist ortion to the down-co nversion process. This analysis shows that the 27.5-GHz LO l eakage should not affect the operation of RX chain, and the frequency plan introduced in the previous section should work. However, measurements will ultimately determine the viability of the frequency plan. A new frequency plan may be needed in case the original plan is found to degrade the RX performance. E.g., a 60GHz LO and its divided-by-4 product (15 GHz) could be used to realize the radar RX and mitigate the LO leakage and image problems. 4.7 Design of Frequency Generation System for the 77-GHz Pulsed Radar Transceiver In the 77-GHz radar transceiver architectur e discussed above, a mm-wave frequency generation system (synthesizer) is a vital circuit block providing high frequency LO signals for Figure 4-16. Block diagram of the 51-GHz PLL frequency synthesizer (in dash box) in the proposed 77-GHz radar transceiver
127 both generating pulses of carrier around 76.5 GHz in the pseu do direct conversion TX and down-converting the received pulses to baseband in the dual conversion RX As shown in Figure 4-16, the frequency synthesizer consists of a 51-GHz PLL, multiple buffers tuned at different frequencies and an up-conversion offset mixer. At the core of synthesize r, the 51-GHz PLL is a 4th-order loop with a charge pump, where the RC loop filter bandwidth is se t to ~250 kHz and the reference frequency is ~225 MHz. With this arrangement, the severe reference spurs shown in mm-wave PLL such as the one in  hopefully can be attenuated to more than 40 dBc. Only divide-by-2 freque ncy dividers  are needed in the PLL, which mitigates the comple xity of mm-wave PLL design. As mentioned in Section 4.5, all circuit blocks are intentionally designed at ~10% higher frequencies than those shown in Figure 4-16 to provide a margin for frequency tuning down due to variations and underestimated para sitic capacitances. In the loop, a 55-GHz VCO with a cross-coup led NMOS pair similar to the 94-GHz VCO discussed in Chapter 3 is utilized. In simulations, the VCO is tuned from 55 to 61 GHz to cover the possible frequency shifting caused by proces s and temperature variations. The operating frequencies of all other components for the synthesizer are tuned higher accordingly in simulations. A chain of asynchronous divide-by-2 static frequenc y dividers provides a division ratio of 256. A 2-stage tapered buffer (Buffer 1) isolates the VCO core from the 1st frequency divider, and Buffers 2 and 3 for driving the LO port of the 1st mixer in the RX chai n and the RF port of the up-conversion offset mixer. ~2m wide differential top metal lines compromising parasitic resistance and capacitance at ~60 GHz are used to distribute the VCO bu ffer output signal to the
128 three circuit blocks. The parasitic inductances of these interconnects are exploited to form series peaking networks to mitigate signal attenuation. The first divide-by-2 circuit plays a critical role in determining the maximum operating frequency of the PLL, and it is the most cha llenging design among the dividers. An injectionlocked frequency divider (ILFD) can be used as the first divider as in [112 ]. ILFD can operate at very high frequencies, but its lo cking range usually is narrow and hence its robustness over process and temperature variations is a concern. From this point of view, the current mode logic (CML) type static divide rs with a wider locking bandwidth and greater tolerance to the variations are preferred. With th e scaling of CMOS technology into nano-scale, MOS transistors are becoming faster and it should be pos sible to use a CML type static divider instead of the ILFD as the first divider stage for the 51-GHz PLL . With the TI 65-nm CMOS process low leakage transistors, a latch-biased static divider modified from  fo r low voltage operation, as shown in Figure 4-17, is designed and optimized to achieve the necessary locking frequency. A Figure 4-17. Latch-biased static frequency divider used for the 1st and 2nd stage of divider. A) The schematic. B) The 1st divider layout with custom metal6 dummy fills and dummy blocks to reduce the parasitic capacitive load for the divider.
129 B Figure 4-17. Continued For this latch-biased divider, th e gate of the latch tail curren t source transistor is connected to a constant bias instead of the input high frequenc y clock signal. In this way, the divider capacitive load to the previous stage is reduced. Th e biased latch also provides regenerative gain for the divided-down signal without the need for waiting to be switched on by the mm-wave clock signal. This decreases the signal transition times and enhances the divider speed. A small sized PMOS transistor is used as a resistive load in the divider because of the relatively large polysilicon resistor varia tion in the TI process. Also, the parasitic capacitive load added by the PMOS load in the TI 65-nm CMOS is smaller and the impact on the maximum locking frequency is small in simulations. Due to the short distances between the back-e nd metal layers and substrate, and the dense metal dummy fills in the TI process, the parasitic capacitances from the wiring within the divider and between divider blocks, and the floating metal dummy fills above the divider are nonnegligible portion of the total capacitive load of divider. For a divider working at high frequencies, these parasitics s hould be considered in the divide r design. Besides, since such parasitics from wirings and dummy fills do not scal e with divider transistor sizes in the first
130 order, to maintain a certain swing at the divi der output and to minimize the RC delay in the positive feedback loop of these flip-flop style static dividers, the transistors need to be sized-up to drive more current into the load. This must be balanced by the fact that the larger the transistor size in the first divider, smaller the input swing set by the preceding VCO buffer. In the implementation, custom metal6 (M6) dummy fills and dummy blocks are juxtaposed with the power supply lines and bypass capacitors (Figure 4-17(B)) to satisfy the metal density rules. This reduces the parasitic loading of di vider by ~8 fF and increases the maximum dividing frequency by ~10%, as shown by the simulated input sensitivity curves in Figure 4-18. A 2-stage tapered buffer (Buffer 4) is inserted between the first two dividers to increase the 1st divider output swing for driving the 2nd divider, and Buffers 5 and 6, hence improving the robustness of divider chain. The 2nd divider in the divider chain probably is as important as the first one for setting the maximum operating frequencies. The input transistors of the 2nd divider contributes a large portion of the capacitive load to the first divider, wh ich then limits the transistor sizing of the 2nd divider. Since the 2nd divider actually works at quite high frequency (~27-30 GHz). Its transistor sizes cannot be too small. Since each divider presents loading to the previous stage, and the cascading of these dividers (DC coupled) changes the bias points of individual divider, it is necessary to co-optimize the VCO and the divide r chain as a whole to make sure the divider chain works properly. In simulations, the first stage latch-biased di vider achieves a locking range between 52 and 71 GHz for a sinusoidal clock i nput with 600 mV peak -to-peak swing and the whole divider chain can successfully divide the VC O output within the tuning range.
131 -6 -5 -4 -3 -2 -1 0 1 40 50 60 70 80 Input Frequency (GHz)Input Power (dBm) w/ Custom Dummy Block w/o Custom Dummy Block Figure 4-18. Simulated input sensitivity plot of the 1st latch-biased static frequency divider. Following the divider chain, a dead-zone free, three-state phase-fre quency detector (PFD) and charge-pump (CP) use the same topology as that in  and . Th e transistor sizes of charge-pump PMOS/NMOS output current sources are optimized to obtain charge/discharge current matching to reduce the reference spurs in simulations. A 3rd order RC loop filter  is used to set the loop bandwidth to ~250 kHz. This along with the rela tively high reference frequency of ~225 MHz further atte nuates the reference spurs. All the tapered buffers (Buffers 1-7) use the balanced common s ource tuned amplifier topology for improved operation at low voltage. Their LC res onant tank loads use spiral inductors optimized with custom dummy metal fills and dummy blocks as discussed in Chapter 2. This reduces the chip area and makes the necessary connections easier. Notice that the 3-dB gain bandwidth of these buffers should not be too narrow. Otherwis e, if the buffers are mistuned,
132 they may adversely affect the whole synthesizer operation. Just like the divider chain, the multistage buffer should be optimized as a whole once the transistor size and inductor load values are roughly decided. Finally, as in the design of amplifier, the stability of these buffers should be maintained. A double balanced Gilbert cell topology, which is the same as the core circuit of the downconversion mixer discussed in Chapter 3, is used in the offset mixer for up-converting the ~55 GHz output from VCO to ~83 GHz. It is designed to maximize the conversion gain for an RF input of 55 GHz and LO input of 27.5 GHz co ming from the VCO output buffer and the first stage divider output buffer, respectively. The simulated mixer convers ion gain is ~ dB at ~83 GHz assuming a 600 mV p-p LO signal at ~2 7-GHz. This LO signal is from the 1st frequency divider output, which can be more easily amplif ied compared to the ~55-GHz signal. A 4-stage tapered buffer (Buffer 7) following the offset mi xer provides the input to the PA in TX. It amplifies the desired ~83-GHz signal by ~8 dB and attenuates the ~27-GHz image and ~55-GHz leakage by more than 30 dB. For single-ended meas urement, one side of the Buffer 7 outputs is terminated by an on-chip 50resistor. The whole up-conversi on mixer chain including all the related buffers consumes ~60 mW in simulations. The frequency synthesizer and receiver chai n of the radar discussed above should be integrated to form a complete mm-wave radar receiver front-end. The de sired LO signals at 55 and 27.5 GHz generated by the synthesizer will need to be delivered with sufficient amplitudes to the two down-conversion mixers in the receiver chain. Multi-stage buffers (Buffer 2 and 3) are needed to drive the LO signals through the relatively long in terconnects to the mixer. In simulations, Buffer 2 has ~9 dB voltage gain at 55 GHz while Buffer 3 has ~ 14 dB voltage gain at 27.5 GHz. Each buffer consumes ~20 mW.
133 4.8 Measurements of the Radar Receiver Fr ont-End Integrated with the Frequency Generation System The receiver chain integrated with the frequency generati on system and the frequency generation system itself discussed in the prev ious sections are implemented using the low leakage transistors of TI 65-nm CMOS process. In this section, we will talk about the measurement of the frequency generation system fi rst, and then that of the receiver chain. Measurement of the frequency generation system: A die photo of the frequency generation system is shown in Figure 4-19. The active circuit area is 0.98X 0.95 mm2 excluding bond-pads. The differential frequency generation ch ip is mounted on a printed circuit board and measured on-chip with GGB W-band and V-band GS/SG pr obes and Agilent 11970W and 11970U harmonic mixers, as shown in Figure 4-20. An HP 8640B signal generator is used to provide the ~215-236 MHz referenc e signal to the PLL. The first frequency divider with Buffer 4, and the offset mixer with Buffer 3 and 7 draw 10 and 54 mA at VDD=1.4 V. The other circuits draw 75 mA at VDD=1.2 V (~60 mA for buffers). The PLL is locked from 55.4-58.2 GHz (low band) and 57.8-60.3 GHz (high band) with the VCO bias current (including 1st stage of Buffer 1) set at 13 and 10 mA. The locking range is lim ited by the VCO tuning range and charge pump supply voltage. Figure 4-21 plots the output fre quencies and single-ended output power of VCO and offset mixer in the lower tuning range. The locked output spectrum and phase noise at 55.5 GHz measured with an Agilent E4448A spectrum anal yzer are shown in Figures 4-22 and 4-23. The phase noise is and dBc/Hz at 0.25, 1 and 10 MHz offsets. Figure 4-24 plots the spectrum of corresponding 83.3-GHz out put of the offset mixer. Its phase noise is dBc/Hz at 0.25-MHz offset. The phase noise at higher offsets could not be reliably measured due to the measurement set-up loss of ~55 dB. The output po wer levels of PLL and offset mixer are ~
134 and dBm at ~55 and ~83 GHz respectively. The corresponding differe ntial output power should be ~ and dBm. Figure 4-25 shows th at in the entire locking range, the PLL output shows phase noise varies from to dBc/Hz at 1 MHz offset and the reference spur varies from -40 to -35 dBc. This work demonstrates the feasibility of usi ng low leakage transistors of a low-cost 65-nm CMOS process for manufacturing cell phone chip s to build a wide-loc king range frequency generation system for the proposed 77-GHz rada r transceiver. Compared to  and  in Table 4-8, This fre quency generation system employs an inductor-less static divider capable of dividing at 60 GHz a nd an offset mixer extending the car rier frequency to ~90 GHz to overcome the limitation of high frequency performan ce of the low leakage transistors being used. It achieves a wider output freque ncy range, better phase noise with lower (20-30%) core power consumption at similar frequencies. Tables 49 and 4-10 summarize the power consumption and measured performance of this system. Figure 4-19. 77-GHz radar freque ncy generation system die photo.
135 Figure 4-20. Measurement set-up for the 77-GHz radar frequency generation system. Figure 4-21. Frequency tuning a nd single-ended output pow er of VCO and offset mixer after deembedding the measurement set-up loss. Th e differential output power should be 3dB higher. 50 60 70 80 90 -0.2 0.2 0.6 1 1.4 Vtune (V)Freq (GHz)-40 -30 -20 -10 0Single-ended Power (dBm) VCO output Offset Mixer Output
136 Ref 0dBm Norm Log 10 dB/ LgAv V1 S2 S3 FC A AA ( f ): FTun Swp Mkr1 55.542 473 1 GHz -10.16 dBm Center 55.542 400 GHz Span 4 MHz Res BW 51 kHz VBW 5 Hz #Sweep 6.12 s 1 A Ref 0dBm Norm Log 10 dB/ LgAv V1 S2 S3 FC A AA ( f ): FTun Swp Center 55.545 125 GHz Span 500 MHz Res BW 1 MHz VBW 510 Hz #Sweep 0.76 s 39.80 dBc 40.50 dBc B Figure 4-22. Spectrum of the PLL 55.5 GHz output A) Span of 4 MHz B) Span of 500 MHz
137 Carrier Freq 55.5 GHz DANL off Trig Free Carrier Power -42.21 dBm Mkr1 1.000 MHz Ref -40 dBc/Hz -86.79 dBc/Hz 20 dB/ 1 kHz Frequency Offs et 100 MHz 1 Figure 4-23. Phase noise of the PLL 55.5 GHz output. Ref -10dBm Norm Log 10 dB/ LgAv V1 S2 S3 FC A AA ( f ): FTun Swp Mkr1 83.313 705 1 GHz -18.25 dBm Center 83.313 785 GHz Span 2 MHz Res BW 51 kHz VBW 5 Hz #Sweep 6.12 s 1 Figure 4-24. Up-converted 83.3 GHz spectrum at the offset mixer output corresponding to 55.5 GHz output of the PLL, showing the loop bandwidth of about 250 kHz.
138 -91 -89 -87 -85 -83 -81 5456586062 Locking Freq (GHz)Phase Noise (dBc/Hz) @ 1 MHz Offset-41 -39 -37 -35 -33 -31Ref. Spur (dBc) Low Band High Band Figure 4-25. Phase noise at 1 MHz offset and reference spur. Table 4-8. Performances of CMOS phase-locked loop above 50 GHz with fundamental VCO. mm-wave PLLs     This Work Frequency locking range (GHz) 45.9-50.5 58-60.4 73.4-73.72 61.1-63.1 55.4-60.3,83.190.4 Supply (V) 1.5/1.2 1.2 1.45 1.2 1.4/1.2 PLL core power (mW) 57 80 88 (without buffers) 78 46 Phase noise (dBc/Hz) @ frequency offset (Hz) At 46 GHz -63.5@100k -72@1M -99@10M At 60.4 GHz -85.1@1M -95 @ 2M -100.4 @10M At 73.5 GHz -88@100 k -108 @ 10M (Estimated by author) At 61.4 GHz -72@100k -80 @ 1M -110@ 10 M (Estimated from plot) At 55.5 GHz -65 @ 250 k -87 @ 1M -112@ 10 M At 83.1 GHz -63 @ 250 k
139 Table 4-8. Continued Ref spurs -40 ~ -27 -53~ -50 <-72 (Estimated by author) -49 @61.4 GHz -40~-35 Differential output power (dBm) -10 @ 46 GHz N/A -27 @ 73.5 GHz -7 @ 61.4 GHz -7.5 @55.5 GHz -15.2 @ 83.1 GHz Technology 130 nm 90 nm 90 nm 90 nm 65 nm (low leakage transistors) Area (mm2) 1.16 X 0.75 0.95 X 1 1 X0.8 1.2X0.7(core) 1.75 X 1.31 0.98X0.95 (core) Table 4-9. Summary of pow er consumption of the frequency generation system Circuit block Supply voltage (V) Supply current (mA) DC power (mW) Only PLL core operating VCO & 1st stage of Buffer 1 1.2 13(low band) /10(high band) 15.6 (low band) /12(high band) 2nd stage of Buffer 1 1.2 3 3.6 Buffer 2 1.2 8 9.6 1st divider 1.4 5 7 Buffer 4 1.4 5 7 2nd divider 1.2 1 1.2 Other dividers 1.2 1 1.2 PFD, charge pump & inverters 1.2 1 1.2 PLL core total 1.4/1.2 25 46.4/42.8 Whole frequency generation system operating Offset up-conversion mixer 1.4 7 9.8 Buffer 3 & 7 1.4 47 65.8 Buffer 5 1.2 18 21.6
140 Table 4-9. Continued Buffer 6 1.2 18 21.6 2nd stage of Buffer 1 1.2 6 7.2 Buffer 4 1.4 8 11.2 Frequency generation system total 1.4/1.2 133 173/170 Table 4-10. Frequency gene ration performance summary Features Specifications PLL locking range (GHz) 55.4-58.2(low band) 57.8-60.3(high band) 83.1-90.4(offset mixer output) Loop bandwidth (MHz) 0.25 Division ratio 256 Phase noise (dBc/Hz) @ frequency offs et -87~ -84.2 @ 1 MHz (low band) -83.3 ~ -80.1 @ 1 MHz (high band) At 83.1 GHz -62.9 @ 250 k Reference spurs (dBc) -40~ -37 (low band) -38.3~ -35 (high band) Differential output power (dBm) -9.1~ -7.1 (low band) -13.8~ -10.9 (high band) -14.9~-20.4 (offset mixer output) PLL core power consumption (mW) 46.4 (low band)/42.8(high band) Frequency generation system power consumption (mW) 173 (low band)/170 (high band) Technology TI 65-nm low leakage bulk CMOS Chip area (mm) 1.75 X 1.31 0.98X0.95 (core) Measurement of the receiver ch ain integrated with frequency generation system: A die photo of the receiver chain in tegrated with the frequency ge neration system is shown in Figure 4-26, which occupies a total area of 2.55 X1.38 mm2 including bond pads. The receiver
141 chain occupies 1 X1.38 mm2 Figures 4-26 (A) and (B ) are the photographs of the whole chip and the zoom-in for the receiver chain portion respec tively. The differential interconnect lines are ~ 250 m and ~ 760 m long between the frequency generatio n system LO buffers (Buffer 2 and 3) and the 1st and 2nd down-conversion mixer LO inputs resp ectively. The width and spacings of these lines are optimized to have ~1 dB propagation loss fo r differential LO signals at corresponding frequencies in HFSS simulations. In the circuit design, these lines are modeled as multiple lumped line inductors with the 2model introduced in Chapter 2. With measurement set-ups similar to that di scussed in Chapter 3 for mixer measurements (Figure 4-27), the receiver chain is characte rized on chip with a Wband GSG probe providing RX input signal, a hybrid probe with DC and AC pins with AC pins landed on circuit baseband (BB) output pads, and a 40-GHz GS probe for connecting an HP 8640B signal generator to provide reference clock for th e frequency generation system. A Figure 4-26. Die photos of the W-band receiver front-end. A) Receiver front-end. B) Receiver chain close-in.
142 B Figure 4-26. Continued Figure 4-27. Measurement setup for the receiver front-end.
143 The receiver input matching characteristic is measured with an Agilent PNA E8361A with two N5260A frequency extension modules. The measurements and simulations of the matching are close as shown in Figure 4-28.The measured |S11| centered around 83 GHz is less than dB from 78.8 to 88.3 GHz and less than dB from 77 to 90 GHz. -25 -20 -15 -10 -5 0 707580859095 Measured |S | Simulated |S |11 11 Figure 4-28. Measured and simulated input matching |S11| of the receiver front-end. The conversion gain of receiver is charac terized with an Ag ilent E4448A spectrum analyzer single-endedly with one differential output terminated w ith a bias-tee and an external 50load. Not having a low frequency external ba lun down to ~MHz range and the concern for phase mismatching of the flexible soft cable att ached to the hybrid probe AC pins make this necessary. VDD of 1.2 V is used for the receiver chain circuit blocks expect for the RF and IF amplifiers that use 1.4 V VDD for better performance. To enhan ce the LO swing at the mixer, 1.5V VDD is used for the two corresponding LO buffe rs (Buffer 2 and 5). Figure 4-29 shows the differential conversion gain (a dding 3 dB to the single-ende d measurement results) of the receiver front-end for an input frequenc y range of ~78-88 GHz with a 55.5-GHz 1st LO signal and a 27.75-GHz 2nd LO signal generated by the on-chip frequency generation system. The
144 simulated conversion gain with 55.5-GHz 1st LO signal and a 27.75-GHz 2nd LO signal in the original design is also shown for comparison. At the baseba nd, the down-converted signal frequency is from 50 MHz to 5 GHz. The receiv er SSB NF is measured from 50 MHz to 3 GHz at the baseband with the method described in Sec tion 3.4. This NF measurement range is limited by the bandwidth of external baseband low noise amplifier in the measurement set-up. The NF data are shifted up from baseband to the 83.25 GHz for easy comparison with simulated SSB NF. -16 -8 0 8 16 76.578.580.582.584.586.588.5 Frequency (GHz)Conv. Gain (dB)8 16 24 32 40SSB NF(dB) Measured Results Simulated Results Figure 4-29. Measured and simu lated conversion gain and SSB NF of the receiver front-end. As can be seen, the receiver chain achieves maximum conversion gain of ~7 dB and minimum SSB NF of ~13.8 dB at 84 GHz RF i nput under a total power consumption of 187 mW (90 mW for the receiver chain a nd 97 mW for the frequency gene ration system). The 3-dB gain bandwidth is ~5 GHz from 80.5 to 85.5 GHz. Notice that in the proposed pulsed radar appl ication, the received signal should be a DSB signal around 83 GHz. Thus DSB NF of the receiver should be used for the communication link calculation. Theoretically, DSB NF is 3 dB lower th an SSB NF if the receiv er conversion gain is the same for the two sideband of the signal. As can be seen in Figure 4-29, within ~ 1 GHz
145 bandwidth around 83 GHz, the receiver conversion ga in has less than 0.6 dB variation. This indicates it is valid to use DSB NF that is 3 dB lower than SSB NF in this range. Hence the minimum DSB NF should be ~ 10.8 dB. For IIP3 measurement, two test tones at 84 and 84.01 GHz are generated using two N5260A extension modules and combined with a DualComm W-band magic-tee. Figure 4-30 shows the measured IIP3 and IP1dB of the receiver are -12 dB m and -21 dBm. Table 4-11 summarizes the detailed power consump tion of this 83-GHz radar receiver. -95 -75 -55 -35 -15 5 -40-30-20-10 RFInputPowerof84and84.01GHz(dBm) BB O u t pu t P ower (dB m ) 750 MHz BB Output 760 MHz BB OutputIP -21 dBm 1dB IIP -12 dBm 3 Figure 4-30. Measured IIP3 and IP1dB of the receiver front-end. Table 4-11. Summary of power consum ption of the 83-GHz receiver front-end Circuit block Voltage supply (V) S upply current (mA) DC power (mW) RF amplifier 1.4 35 49 RF mixer 1.2 5 6 IF amplifier 1.4 20 28 BB buffer 1.2 7 8.4 Whole RX chain 1.4/1.2 68 89.6
146 Table 4-11. Continued PLL core w/o LO buffers and offset mixer chain 1.4/1.2 17 36.8 RX LO buffers (Buffer 2 and 5) 1.5 40 60 Operating RX w/ onchip LOs 1.5/1.4/1.2 133 187 In summary, the receiver is measured to have input matching better than 8 dB from 77 to 90 GHz. Its peak conversion gain is ~ 7 dB ar ound 83 GHz, and the lowest SSB NF is ~ 13.8 dB. Its IIP3 and IP1dB are -12 dBm and -21 dBm. The measured peak conversion gain and lowest NF of the receiver ch ain are ~8 dB lower and ~3.3 dB higher respectively than the simulated results. Th rough simulations, the differences are mainly attributed to two factors assuming the RF down-conve rter gain is close to the designed value based on previous mm-wave ci rcuit design and measurement experience. First, on-chip LO swing at the mixer is smaller than the 600 mV p-p as designed. According to simulations, the over-estimated indu ctor Q, parasitic sour ce degeneration inductor and dummy metal loadings to th e inter-stage interconnects in th e LO buffer layout reduces the on-chip LO swing by at least 6dB at both 55 GHz and 27.5 GHz. Such a reduction of LO signal swing at the LO ports of the 1st and 2nd down-conversion mixer degrades their peak conversion gain by ~2.5 dB and ~1.5 dB, respectively. Second, there are ~ 300 m interconnect lines connecting the 2nd mixer output to the high pass filter in the circuit la yout, which forms a ~200-fF capaci tive load to the IF cascode amplifier. This is not accounted for in the original design. After adding this parasitic capacitance in the simulation, the cascode amplifier gain drops by more than 3 dB.
147 The gain reduction of the above circuit bloc ks increases the receive r NF by ~1.5 dB in simulations. The rest of the differences of conversion gain and NF are attributed to the on-chip ground parasitics and component mo del inaccuracy as discussed in the 80-GHz down-converter design in Section 4.4. The above analysis points out the direction of improving the receiver performance, which includes enhancing LO drive and re tuning circuit blocks with more careful attention to the layout parasitics and more accurate models. 4.9 Summary In this chapter, a mm-wave pulsed radar syst em operating at 76-77 GHz is proposed for collision avoidance and adaptive cruise control in vehicles. The mm-wave receiver circuits of radar transceiver are desi gned in the TI 65-nm CMOS process to demonstrate the feasibility of using low-cost digital bulk CMOS to build compact mm-wave automobile radar modules affordable by average consumers. First, an 80GHz down converter consisting of a single-ended multi-stage tuned amplifier and a single-ended pseudo-differential down-conversion mixer is implemented. No on-chip balun is needed on the received signal path and the single-ended to differential conversion is done in the mixer. Th e down-converter obtains ~16 dB conversion gain and ~ 9 dB NF in measurements. A dual conversion heterodyne arch itecture for the rada r transceiver is pr oposed, where two on-chip dipole antennas will be used as mm-wave f eeds to an external high gain antenna. After that, the designs of 77-GHz receiver chain a nd the accompanying frequency generation system based on a 51-GHz PLL are presented. The test st ructures of frequency generation system and the receiver front-end are implemen ted with low leakage transistors in the low-cost 65-nm bulk CMOS process measured once again.
148 The frequency generation system employs an inductor-less static divider capable of dividing up to 60 GHz and an offset mixer exte nding the carrier frequency to ~90 GHz. Phase noise of dBc/Hz at 1 MHz offset for 55 GHz carrier and 63 dBc/Hz at 250 kHz offset for 83 GHz carrier are measured. The differential output powers are a nd dBm for the two carriers, respectively. The system consumes ~170 mW power. A heterodyne receiver front-e nd including the receiver chai n and frequency generation system is successfully demonstrated. It achieve s ~7 dB peak conversi on gain and ~13.8 dB lowest NF in the input frequency range of 8386 GHz with LOs being generated on-chip. Its IP1dB and IIP3 are and dBm. The receiv er consumes ~190 mW power. The circuit works in this chapter suggests the feasibility of using low leakage transistors of a low-cost 65-nm CMOS process to build a co mplete receiver front-e nd integrating a widelocking range frequency genera tion system for the proposed 77-GHz radar transceiver.
149 CHAPTER 5 SUMMARY AND FUTURE WORK 5.1 Summary Millimeter-wave receiver front -end circuits in digital bulk CMOS and their applications were studied. It is recognized that with CM OS scaling and device performance improvements, mm-wave circuits fabricated in CMOS are po ssible and this can pr ovide a low-cost, highintegration alternative to thos e based on other IC technologies. Some general trends for CMOS technology as it is scaled down to nano-meter re gime are reviewed. Based on these, circuit design challenges including keeping the circui t simple and dealing with low voltage are identified. A low-cost 65-nm digital bulk CMOS technology with low leakage MOSFETs offered by TI is used as the platform to dem onstrate the feasibility of building mm-wave frontend circuits. Essential active devi ces and passive devices in this process including their models are discussed. An 80-GHz tuned amplifier, a 94-GHz VC O and a 77-GHz down-conversion mixer are designed and experimentally demonstrated. N on-ideal on-chip ground effect and substrate coupling effect are examined, and critical mm-wave measurement set-ups are described. Measurements show that the 80-GHz amplifier is able to reach 12 dB gain at 79 GHz and ~10 dB NF. Its IIP3 and IP1dB are ~-21 and ~-11dBm. The 94-GHz VCO achieves dBc/Hz phase noise at 1 MHz offset and 5.8% tuning range. Af ter deembeding the on-chi p RF input balun loss, the 77-GHz mixer shows ~.5 dB conversion gain at 76.3 GHz and NF of ~12 dB between 7677 GHz. Its IIP3 and IP1dB are ~-4 and ~-13dBm. Based on the performances of these mm-wave circuit bloc ks, two applications are considered including a mm-wave wi reless interconnect syst em to alleviate issu es of conventional wire-based interconnect system such as complex I/O wiring, larger package area and limited data
150 bandwidth, and a 77-GHz automobile radar sy stem. The wireless interconnects system architecture, multiple access, duplex and modula tion scheme, radio circuit architecture and associated frequency plan are explored and presented in the appendix. The 77-GHz radar application has been chosen as the focus of this research. A 77-GHz pulsed radar system for automobile is proposed and its radio architecture is presented. The radio integrates a transmitter and a receiver on one chip, with two on-chip dipole antennas for mmwave feeds for an external antenna. As the firs t step towards a more complete 77-GHz radar receiver front-end in the lo w-cost TI 65-nm CMOS technology, an 80-GHz down-converter consisting of a tuned amplifier and a mixer is demonstrated. The down-c onverter shows ~16-dB conversion gain and 9-dB NF at 80 GHz, and IIP3 and IP1dB of ~-16 and dBm. Next, a receiver chain and a frequency ge neration system for the radar system are demonstrated in the same technology. Measurements show that the frequency ge neration system achieves phase noise of dBc/Hz at 1 MHz offset at 55 GHz and -63 dBc/Hz at 250 kHz offset at 83 GHz. Its differential output power is and dBm at the two out put frequencies while consuming ~170 mW power. The heterodyne receiver front-end integr ating the receiver chain and frequency generation system achieves ~7 dB peak conversion gain and ~13.8 dB minimum SSB NF (~10.8 dB minimum DSB NF) in the input frequency range of 83-86 GHz with LOs being generated on-chip. Its IP1dB and IIP3 are and dBm respectively while consuming ~190 mW power excluding that for the offset mi xer and related buffers in the frequency generation system. The successful demonstration of these two circuits fulfils the goal of this PhD research work, demonstrating the feasibility of using low-co st low leakage CMOS technology to realize a complete receiver front-end for 77GHz auto mobile radar systems (and other mm-wave
151 applications). It is a significant step towards a complete mm-wave radar transceiver in such lowcost technologies. Although the goal of this resear ch is successfully met, addi tional efforts are needed to improve the receiver performance and enhance the robustness of th e circuit and system towards a real low-cost product. 5.2 Future Work First, to improve the 77-GHz receiver front-e nd performance, a better LO buffer design is needed to increase LO signal swi ng, together with re-optimized interconnect lines and floor plan. The IF cascode amplifier will be better frequency tuned. On-chip antennas will be integrated and co-optimized with RX. The targ eted radar module including ch ip, package, PCB cavity, the waveguide transitions and external high-gain antennas should be designed and demonstrated. Second, improved wide-band modeling includin g not only at the fundamental frequency band but also in sub-harmonic and harmonic fre quency bands is needed for passive components such as inductors, capacitors, transmission lines, ground and even the passive parts of transistors and varactors. Equivalent circuit models should not only be compared to 3-D EM simulation results but also to the actual measur ement results of test structures. Also, it is critical to develop a rapid, user-friendly design flow that can verify the physical layout s of the planar structures at such high frequencies against their models used in simulations to ensure the consistency between the design and implementation. In addition, alternative structures and layout techniques of onchip components should be explored for better performance, e.g., using flux capacitor instead of metal plate capacitor to obtain higher Q. In the long run, alternate frequency plans and associated frequency ge neration circuits and systems should be explored. Baseband amplifier and filter will be added to enhance the signal strength and quality to make the signal acceptable for ADC circuits following the receiver front-
152 end. A narrow-pulse-generator circuit and pulse-modulated mm-w ave power amplifiers should be studied and integrated with the receiver front-end and whole radar module. Also, DC bias generation, self-testing, calibra tion and tuning/compensation circu its, DAC/ADC circuits, digital signal processing and control circui ts should be integrated with th e radar transceiver front-end to fully integrate the radar function. All mm-wave/ RF/analog circuit blocks will be re-optimized together focusing on circuit robustness over pr ocess/temperature/voltage variations, low voltage/low power operation, tunability and reliability.
153 APPENDIX SYSTEM LEVEL STUDY FOR WIRELESS INTER CHIP INTERCONNECT IN MM-W A VE BAND A.1 Interconnect Bottleneck Based on 2006 International Technology Roadma p for Semiconductors (ITRS), at the 35nm technology generation around year 2012, the number of pins in packaged chips will be on the order of 1000-5000. ~35% of these pins (~ 350 to 1750) will be input/output (I/O) pins . At that time, a ball grid array (BGA), which is a popul ar package type for VLSI circuits with many I/Os, is expected to reach the manufacturing precision of ~0 .5 mm pitch among the soldering balls on the package. Hence, the BGA area needed to accommodate all the pins is around 3.5 cm x 3.5 cm. Since each I/O pin requires a wire to connect to other chips/components for data communication in an electronics system consisting of multiple chips/modules, an even larger PC board area is required to accommodate the I/O wires. This is expected to increase the overall cost of an electronic system drastica lly and package/board expense is expected to be dominant in the cost structure. Packaging complexity, package and board area increase and the associ ated rising cost are just some facets of the problems for the trad itional wire-based in ter-chip interconnects. Bandwidth, latency and cross-talk are also se rious issues for wire-based interconnects. The baseband data bandwidth in modern personal computer systems already reached beyond 64-Gbps thanks to the succe ssful scaling of CMOS technology, and this trend is expected to continue at a quick pace. Meanwhile, the associated clock rate for the high-speed data processing is becoming higher and reaches up to several GHz these days. With this rapid increase of data rate and cloc k rate, the traditional interconnect has become the bottleneck for inter-chip communication. The inherent low pass filter response of interconnect channel (even with relays) will ul timately limit the data rate, while the latency
154 experienced by a signal propagating on the wire is bounded by the wave pr opagation speed in the package/board medium for the interconnect system. Moreover, the cross-talk among the ultra dense I/O wires degrades signal integrity and can fail the communication link. Obviously, these problems will become more severe when the digital circuits in future generations of CMOS technology operate at ev en faster speed. Some potential alte rnatives to traditional wire interconnect, including system-in-apackage (SIP) , 3-D integra tion with wires and vias , or with wireless proximity communication technology , optical interconnect  and RF interconnects (on transmission-line) using microw aves , try to address the traditional wire interconnect problem from the basic ideas of reducing interconn ect distances and e nhancing interconnect bandwidth and propagation speed. Table A-1 li sts the pros and cons of these options. Table A-1. Some possible solutions for the wire interconnection bottle-neck Possible solutions for wire interconnection problem Features System-In-A-Package (SIP)  Need new process, new material and research still in conceptual phase 3-D integration, chip stacking, vias through chips , wireless proximity interc onnection  Need additional process st eps such as substrate thinning, chip alignment and adhesive/bonding, heat dissipation problem, reliability and repeatability problem and for very short distance (<20 m) Optical interconnection  Difficult to integrate with silicon CMOS and silicon optics research just started RF interconnection using t-line  Still rely on t-line for interconnection Wireless interconnection proposed In this thesis No wire interconnection needed and data transmitted by EM-wave The SIP technology requires increased complexity and cost of a vari ety of materials and process technologies needed to integrate all the required functions in a package. Heterogeneous integration using 3-D technology ha s to deal with the problems of cost, heat management,
155 robustness and circuit architecture paradigm shift. The chip stacking and proximity communications have the similar heat removal c oncern as the 3-D circu its, besides requiring a robust and inexpensive stacking pr ocess compatible with mainstr eam IC manufacturing process. The interconnect challenge can al so be addressed using optical interconnects. This approach however present the challenges of either inte grate light-emitting optical devices in III-V technology on silicon dies, or develop innovative silicon optical devices. Either way, this technology remains challenging and expensive. Las tly, wireless interconnect, seems to the one that is the most compatible with mainstream CMOS technology and it will be examined more closely. A.2 Wireless Interconnect In a wireless interconnect system, both RF transmitter (TX) and receiver (RX), and radiation structure (such as an on-chip dipole antenna) are inte grated with digital circuits. Baseband high speed signal is modulated onto an EM wave by the transmitter and radiated out to the medium. Then it is picked up by a receiver on the other side of the interconnect and translated back to the original data. One RF transm itter/receiver pair thus functions as one pair of I/O pins in wire interconnect and can be shared by multiple I/O pairs as shown in Figure A-1. A wireless interconnect has the potential advantage that the signal propagating is no longer limited by the package/board medium, since the me dium now can be just air for example. The wire interconnect latency is greatly alleviated and communication sy nchronization could be obtained in fewer clock cycles. Th e interconnect bandwidth is also drastically expanded in the sense that a frequency spectrum far beyond DC-t ens of GHz as used by high-speed wireline serial links can now be empl oyed for data transmission. The interconnect speed bottle-neck is shifte d back from the prorogating medium/channel to the circuits and devices (incl uding antennas) to a certain degr ee. This has very important
156 implications in the big picture: Wireless interconnect bandwidth scales together with CMOS device speed. Package Package Package ChipChipChip Wired Inter-chip Interconnection Wired Inter-chip Interconnection Wired Inter-chip Interconnection Package Chip Wireless Inter-chip Interconnection With On-chip/board Antennas Package Chip Wireless Inter-chip Interconnection With On-chip/board Antennas Smaller package area Less I/O wiring Lower packaging cost Smaller package area Less I/O wiring Lower packaging cost Figure A-1. Wireless interconnects help re duce multi-chip package area and I/O wiring complexity By 2012, with CMOS RF circu its possibly able to work up to 130 GHz (1/3~1/4 of fT), 100 GHz or even larger bandwidth can be used by wireless interconnect circuits . That translates to a data bandwidth of 50-100 Gbps assuming a moderate bandwidth efficiency of 0.51 bit/sec-Hz. Such data rate implies a inter-chip wireless interconnect may be better suited for systems with a moderate data rate and a more stringent requirement for size and cost such as smart mobile devices in consumer applications. Suppose a wireless interconnect system provides 50 Gbps data bandwidth, then it ca n replace 200 I/O pins in traditional wired interconnect with 250 Mbps data rate each (as a reasonable number for the moderate data rate applications). This means a si gnificant reduction of I/O pin c ounts, wires and package/board area. Actually, with an intelligent design of multiple access/multiplexing scheme and communication network protocol supporting dynam ic resource sharing among multiple I/O data paths, the data throughpu t can be increased in a wireless interconnect system, resembling a wireless cellular network. This will further reduce I/O pin counts, etc. In addition, the broadcasting capability and reconfiguration flex ibility inherent with wireless communications
157 are inherited by wireless interconnects. As a si mple example, by just switching RF carrier frequencies, the wireless I/O pairing can be ch anged if frequency division multiple access is used. With many potential advantages said above wireless interconnect s face tough challenges from circuit to system: high data rate (~ Gbps ), low latency, low bit error rate (BER, ~10-16), integration of antennas in a lossy silicon subs trate, management of noise and unwanted signal coupling, and fabrication of low power, robust CM OS high frequency circu its, just name a few. All these challenges must be overcome before a practical wireless interconnects system can be realized. Not all of the questions of such a complex system could be answered by a single dissertation. In the following sect ions, a system-level design for a wireless interconnects system operating in the 60-80 GHz band is described for chip-to-chip communication within a multichip module (MCM). A.3 Wireless Interconnect at 60-80 GHz There are several reasons for studying wire less interconnects at 60-80 GHz instead of lower or higher band in this proposal: First, due to the short wavelengths in this band, small size antenna can be built on-chip or with the package, and the antenna efficiency can be improved. A half-wav e dipole at 60 GHz is only ~1.2 mm long on a silicon die. On-chip passive resonant components on-chip are also be smaller at such high frequencies. Second, for RF tuned circuits, absolute frequency bandwidth is larger at higher carrier frequencies. Even with similar design as those used for narrow band communication in low GHz range, the circuits at 60-80 GHz can easily obtain ~GHz bandwidth, which suits the need of ~Gbps data links. Third, at 60-80 GHz, the wireless interconnects should be less vulnerable to the multiple GHz noise coupled from the baseband high-speed data and clock circuitry. Lastly, the circuits developed in this band coul d demonstrate the feasibility of
158 low-cost mm-wave circuits in digital bulk CMOS and be used for other mmwave applications as well. The wireless interconnect system should be synchronized as is done in systems employing conventional interconnects. A synchroni zation system, as an i ndispensable part in modern digital communication, is itself a broad area with many active researches. An in-depth study is beyond the scope of this proposal. For this re search, a system clock is assumed to be distributed among multiple chips using transmission lines printed on the PC board, which greatly simplifies the radio implementation by eliminat ing the synchronization and frequency offset problems. Next, different aspects of the wireless interconnect system in 60-80 GHz will be discussed by first looking at the system architecture, sp ecification and the wireless channel. Then, link margin will be calculated. After that, radio circ uit architecture, and circuit block specification will be studied. A.4 System Architecture and Specification Before system architecture and specification can be conceived, basic features and design directions of the wireless inte rconnect system are identified. First, the maximum communication distance is assumed to ~10 cm, which is consistent with the general size range of portable electronic s. Second, multiple I/O data paths are to be supported by multiple wireless interconnects, an d a multiple access scheme. Third, aggregate data rate is to be ~ Gbps to support moderate da ta rate devices, either large bandwidth or high bandwidth efficiency modulation is needed. Fourth, BER is to be ~10-16, this requires relatively high power from transmitter, a relatively low loss signal propagation path and high sensitivity on the receiver side. Fifth, low latency, low power, small area and small circuit overhead are
159 desired. These point to a simple system and circuit architecture. Lastly, a way to integrate onchip/package/board antenna is needed. For multiple access schemes, generally there are choices of FDMA, CDMA and TDMA. FDMA is the most straightforward, but a multip le carrier generation method should be found in mm-wave band and careful freque ncy planning is needed to a void mutual interference and linearity problem from multiple TX/RX pairs in an FDMA system. Simple TDMA can be implemented with MUX and DEMUX circuits, while simple CDMA using orthogonal codes needs spreader and de-spreader. They both can be used with frequency division multiplexing (FDM) to increase the number of I/O data pa ths sharing one frequency channel (and hence remove more I/O pins). In a ddition, CDMA provides the potential flexibility of switching I/O pairs by just changing the spreading code of a data path with software. The spreading to a certain degree also can alleviate possibl e multi-path effect in the cha nnel. However, both TDMA and CDMA add latency, circuit overheads and synchronization co mplexity to the system. Here, FDMA is chosen for furt her exploration because of its simplicity and effectiveness. Besides, TDMA+FDM and CDMA+FDM could bot h benefit from the FDMA study since their FDM portion has no difference virtually from FDMA Figure A-2 illustrates an FDMA wireless interconnect system, its simple ra dio architecture can be seen. A wireless interconnect can be ei ther uni-directional or bi-directional, just as traditional wire interconnects. There are mainly two types of duplex schemes for bi-directional communication: frequency divi sion duplex (FDD) and time divi sion duplex (TDD). The main difference between these two is that FDD uses two frequency bands for transmitting and receiving, while TDD transmits and receives in di fferent time slots with TX/RX sharing the same frequency band. What should be noted is that these two can actually be used together to achieve
160 better isolation among multiple TX/RX I/O pairs co-existing on the same chip. Besides both schemes do not have to share the same an tenna, although that may cause larger area consumption. For the time being, TDD is selected as the duplex scheme for the system due to its relatively easy frequency pla nning, ability to integrate TX /RX with one T/R switch, and relatively high spectrum efficiency. However as will be seen later, TDD alone may not be sufficient due to the mutual coupl ing among antennas in the system. Figure A-2. FDMA wireless inter-chip interconnec tion system architecture When it comes to the choice of modulati on scheme, although a higher order modulation could reach higher bandwi dth efficiency, their SNR requirements are higher to achieve the same BER and their circuit overheads are larger compar ed to the lower order ones. As previously described, large bandwidth resour ces are available at mm-wave frequencies and simple circuits and system architecture are pref erred. With these, several lo w order modulation methods have been examined; some of their salient features and circuit im plementation considerations are summarized in Table A-2.
161 Table A-2. Comparison of seve ral low order modulation schemes Modulation scheme Bandwidth efficiency (Rb/RFBW) (bps/Hz) Power efficiency (Eb/N0) (dB)@ Pe=10-6 Receiver detection coherent/noncoherent TX power amplifier (PA) linearity requirement Minimum ADC # of bits required BPSK 0.5 10.5 Simple/DPSK Moderate 1-bit CPBFSK (orthogonal) 0.4 13.5 Multiple mixer/ matched filters Low 1-bit BFSK (nonorthogonal) 0.5 ~21 Multiple mixer/ matched filters Low 1-bit OOK 0.5 13.5 Simple/envelope detector Low 1-bit GMSK ~1 for BT = 0.5 ~12 for BT = 0.5 I/Q demodulator/ frequency discriminator Low 1-bit Examinations of the modulation schemes in the tables show that BPSK has the advantages of low SNR requirement, simplicity and compatib ility with FDMA architecture. By simply adding differential encoder and d ecoder, BPSK signal could be de tected differentially (DPSK), which adds some immunity to imperfect s ynchronization and phase offset between the communication pair. In the case where high phase noise is presenting the local oscillator (LO) signals for up/down-conversion such as this mm -wave wireless interconnect application. DPSK should perform better than BPSK. Interestingly, for the same low BER (high SNR) as 10-16, the minimum SNR requirement of DPSK is almost equal to that of BPSK, in contrast to the common im pression that incoherent detection must need ~3 dB higher SNR for the sa me BER compared to coherent detection. This gives a decisive edge for selecting DPSK as th e modulation scheme for th e wireless interconnect system. As discussed in Chapter 2, it is difficult to obtain high output powe r from a single mmwave power amplifier in CMOS technology. This limits the achievable output power from the
162 transmitter of wireless interconnects. Thus an an tenna with higher efficiency and low loss EM wave propagation path are highly desirable to enhance the received power level and link quality. This will be presented in the next section. A.5 Low Loss Propagation Path and Efficient Antenna in a Package As shown in Figure A-3, multiple chips/co mponents are assembled within a multi-chip module (MCM), which then can be incorporated into a larger electronic system. Within this package, chips can communicate with each othe r through wireless interconnects. EM wave radiated out of antennas propagates between the top of package and the bottom multi-layer substrate. Such a propagation environment is a fixed channel. But, obvi ously there will be multiple reflections and scattering from all th e components, hence multi-path fading and time spreading effects from the channe l are still present. Received si gnal strength could be harmed and inter-symbol-interference (ISI) could rise to the degree intolerable to ~Gbps data. Besides, mutual interference among chips could be seri ous. Moreover, path loss of mm-wave signals propagating in the package is expected to be high even at such short distances because of the high frequencies. Figure A-3. Multiple chips/components in an MCM
163 To alleviate such issues wit hout resorting to a complex rake receiver or equalizers, the propagation environment may be modified in reasonable ways that do not require dramatic changes. Figure A-4 il lustrates such a way. Figure A-4. Enhanced wave pr opagation environment for an MCM Conceptually, metal trace can be placed on a multi-layer su bstrate inside the MCM under the chips, and use the top conductiv e package cover to form basica lly a parallel plate waveguide to reduce EM wave propagation loss. Actually, meta l shields or enclosures will need to be added anyways to avoid any spectral usage violation of re gulations from FCC or other agencies. Meanwhile wave absorbing material, e.g., lossy silicon slabs can be f illed in the package alongside the bottom metal trace to reduce wave EM reflections, multi-path effects and unwanted coupling. Such modifications should be relatively easy and inexpensive. Multiple structures capable of radiation are explored as antennas for the wireless interconnects. Among them, on-chip dipole and slot antennas mainly radi ate electric field (Efield) in the horizontal direction and have relatively weak vertic al E-field, thus they are not optimal for use in the proposed parallel-plate waveguide-type ch annel. Flip-chip bumps for the chip packaging have very low efficiency for the targeted 60-80 GHz band. Bond wires for the chip, as could be obtained from standard low-co st packaging process, actually behave like a monopole and have stronger vertical E-field . HFSS simulati ons show that an ~ 1-mm long
164 bond wire could be an effective radiator for th e proposed waveguide cha nnel at 60-80 GHz. As an example, in Figure A-5 , simulated |S11| for a bond wire of about 1-mm long shows that it is well matched around 60 GHz. Fre q uenc y ( GHz ) 585960616263 5764 -26 -24 -22 -20 -18 -16 -14 -12 -28 -10 |S 11 | (dB) Figure A-5. Simulated input return loss of a bond wire antenna around 60 GHz . A series of test structures consisting of silicon chips, gold bond wires and bond pads on FR-4 board is fabricated as shown in Figure A-6 . Figure A-6. Bond wire antenna test structures consisting of silicon chips, gold bond wires and bond pads on FR-4 board . The antenna pair gains for a pair of bond wi res separated by a vari ety of distances and angles are characterized with or without a me tal cover on them. The measurement (Figure A-7 ) show that bond wire antennas are suitable for use at ~ 60 GHz and have antenna pair gain
165 (Ga) of ~ dB at 10-cm separation with a me tal cover placed 2 mm above the PCB. With the efficiency degradation due to the match variatio ns and coupling loss to nearby metal structures such as other bond wires, Ga is found to be ~ dB at 10-cm separation in the worst case. Finally, it is observed that adding metal cover above the FR-4 board did increase the antenna pair gain by ~10 dB. 1 2 3 4 5 6 7 8 9 10 -60 -55 -50 -45 -40 -35 -30 -25 Distance (cm) Ga (dB) w/o cover w/ cover at 2-mm height Friis formula (path loss) Figure A-7. Measured antenna pair gains of bond wire antennas with and without metal cover . A.6 Initial System Specification a nd Simple Link Budget Analysis With the above discussions on system architect ure, modulation and antenna pair gain, we reach an initial system specifi cation (Table A-3) for the FDMA wireless interconnects in the MCM. The channel frequency spac ing is set to 3 GHz, which satisfies the spectrum need of 1Gbps DPSK signal per channel and leaves 1 GHz guard band between ch annels to alleviate adjacent channel interference. This frequency plan facilitates a compact multiple LO generation scheme that will be discussed later. The link budg et for the wireless interconnect with bond wire antenna is calculated from E quation A-1 and A-2 based on data in Table A-3 and A-4. The results are shown in Table A-4.
166 Table A-3. System specification for the frequency division multiple access wireless interconnect system around 60 GHz Features Specifications Data I/O path # 3 Multiple access FDMA FDMA channel # 3 Data rate/ channel (Gbps) 1 Modulation DPSK Duplex TDD RF BW/ channel (GHz) 2 Channel spacing (GHz) 3 Channel freq (GHz) 60,63,66 Bit error rate 10-16 SNR required (dB) 13.5 Total data rate (Gbps) 3 RX Sensitivity (dBm) = -174(dBm) + RX NF (dB) + SNR Req(dB) + RF BW(dB) (A-1) Link Margin (dB) = TX Power(dBm) + Ant. Pair Gain(dB) RX Sensitivity(dBm) (A-2) Table A-4. Link budget analysis for one wireless I/O path for a FDMA wireless interconnect system operating around 60 GHz Features Specifications Channel bandwidth (RF BW) (GHz) 2 Bit error rate (BER) 10-16 Modulation scheme DPSK SNR required (dB) 13.5 Background noise level (dBm/Hz) -174 Transmitter (TX) output power (dBm) 3 Wireless channel length (cm) 10 Antenna pair gain (dB) (best/worst) ~ -41/-52 Receiver sensitivity (dBm) -58 Receiver noise figure (NF) (dB) 10 Link margin (dB) (best/worst) ~19.5/8.5
167 In the link budget analysis, a re ceiver with ~10 dB NF and a pow er amplifier (PA) with ~3 dBm linear output power around 60 GHz are assumed based on th e amplifier work in Chapter 3 These are reasonable performance target for circu its that can be implemented using TI 65-nm CMOS. A 9-dB link margin in Table A-4 is obtained with the wors t bond wire antenna pair gain, most likely the antenna pair gain will be several dBs better than -52 dB. This will give additional link margin, which hopefully would be ab le to cover the SNR degradation from some practical implementation factors like the in sertion loss of T/R switch, adjacent channel interference, noise contribution from stages foll owing the RX front-end, phase noise of RX LO, and additional antenna coupling lo ss etc. To more carefully look at these issues, the radio circuit architecture will be discussed next. A.7 Radio Architecture Direct conversion (homodyne ) is selected as the radio circ uit architecture for the wireless interconnect system proposed here due to seve ral advantages over dual conversion (heterodyne) for this specific application. First of all, it needs fewer circuit bloc ks to obtain high data rate communication, potentially making the TX/RX circ uit and the whole in terconnect system simpler and consuming less power and silicon area. The image problem is eliminated without resorting to a mm-wave image rejection filter and mixer, and IF filter. The LNA and mixer in the receiver can cover the targeted 8-GHz band around 60 GHz relatively ea sily; Baseband LPF and amplifier acting as a channel filter together, can be constructed and handle the 1 Gbps data. The multiple carrier generation scheme is simpler in direct conversi on because only three LOs are needed. They can be shared for the TX and RX mixers (in th e TDD scheme) corresponding to the three FDMA
168 channels. I/O data path switching could be accomplished by switching LOs among different TX/RX pairs. The three FDMA channels in the sy stem can share the same mm-wave circuits. Use of direction conversion architecture does present design challenges. PA pulling in the transmitter can be resolved by having a frequency offset between the VCO output frequency and the PA operating frequency in the carrier gene ration scheme . Excessive low frequency flicker noise and DC offset can be filtered ou t by using a high pass f ilter or a DC offset cancellation loop. Figure A-8 shows block diagrams for a singl e direct conversion transceiver with TDD scheme that can be used for the wireless interc onnect. At the baseband, a simple RC HPF after the LPF and amplifier combination can remove th e low frequency (below several MHz) flicker noise and DC offset (from both the circuits and LO self-mixing), before the 1 Gbps signal entering the final amplifier stage and being samp led by ADC. Since the data bandwidth is wide, the 3-dB BW of HPF can be set up to several MHz without degrading th e signal quality much. LO leakage/radiation problem should be manageable based on the high isolation measured in the tuned amplifier block in Chapter 3. A phase-locked loop (PLL) freque ncy synthesizer needed to implement the FDMA system is presented in Figure A-8. Key considerations for this synthesi zer architecture will be explained briefly in the following paragraphs due to the vital importance of this circuit for the success of the entire wireless FDMA interconnect system. As shown, the frequency synt hesizer architecture allows th e simultaneous generation of three needed carriers at 60, 63 and 66 GHz in one phase locked loop (PLL), avoiding multiple VCOs and multiple loops on one die, which in turn lowers area and power consumption, and eliminate the potential mutual inte rference problem among the VCOs.
169 In this synthesizer, there is no need for a low power efficiency and area consuming frequency tripler or doubler as in , , and the poten tial harmonic pulling of VCO is avoided. LO frequencies are placed at least 6 GHz away from the VCO frequency by a minimal number of offset mixers. This mitigates the PA pulling problem is mitigated. The relatively large offset frequency also makes it easier to remove the image frequency generated in the mixing process without using a single-si de-band (SSB) mixer . A B Figure A-8. Radio circuit architectures: A) Direct-conversion transceiver architecture with T/R switch that can be used by the FDMA wireless interconnect operating around 60 GHz. B) 60 GHz band frequency synthesizer architecture for the transceiver.
170 The buffers before and after the mixers coul d amplify desired signals, isolate different stages, and help filter out unwanted images, harmonics and intermodulation products that may arise. Notch filters can be inserted to attenu ate images and some potentially strong spurs around 60-GHz band more. Such as the mixing products of the 54 GHz from VCO and the harmonics of 6, 9 and 12 GHz signals. In practice, mm-wave no tch filters may require multiple test runs to properly tune the frequency righ t and their insertion loss may not be low. The bandpass response of multi-stage tuned buffers shoul d provide sufficient suppression. Through careful frequency planning, only di vide-by-2 and divide -by-1.5 frequency dividers are needed in the PLL, which eases the CMOS divider design at high frequencies. It is preferable that all the dividers are implemented as a flip-flop style static divider that has wider locking range than a dynamic divider, and can potentially lower area a nd power consumption. However, the CMOS technology available will ultimately determine whether mmwave/microwave static dividers can be used. In this respect, the first di vide-by-2 circuit and the two following divide-by-1.5 dividers are expected to be the bottlenecks of this loop design. High frequency dynamic divider remains an op tion for these critical dividers here. High data rate and low BER wireless communi cation system as the one discussed here requires low phase noise and low spur power levels in the LO signals. Advanced design system (ADS) simulations shows that an LO with -85 dB c/Hz phase noise @ 1 MHz offset is needed at the RX to achieve the 1 Gbps BPSK data link without degrading BER. The phase noise of CMOS mm-wave VCO has always been a concern . Fortunately, the 94-GHz VCO design presented in Chapter 3 shed some light on the feasibility of satisfying the phase noise requirem ent using the TI 65-nm process. In addition, a PLL surely can help suppress VCO in-band (loop bandwidth) phase noise.
171 Generally speaking, in a simple integer-N PLL, reference spurs suppression can be increased by choosing a smaller loop filter (LF) bandwidth and lower division ratio, but the VCO in-band phase noise suppression is degraded and LF area is increased. Lastly, the final carriers output from the synthesizer will suffer from additional phase noi se and thermal noise from the mixing, amplifying and filtering. A.8 Duplex Scheme and System Architecture Reconsideration Due to the advantage of easy frequency planni ng, ability to integrate TX/RX on the same carrier frequency with one T/R switch, and potentia l spectrum usage efficiency increase, TDD is initially selected for the wire less interconnect system. Incidentally, a TDD system does not necessarily need to use a T/ R switch and share one antenna. The TX/RX time slot switching could also be accomplished by other means, such as controlling the TX PA and RX LNA power supply on/off. Figure A-9. Bilateral wireless FDMA interconnect in TDD scheme is realized with three RF transceivers and bond wire antennas on each chip
172 To implement bidirectional communication in the three FDMA channels as planned, three TXs/RXs should be integrated on one die, whic h are challenging. One possible TDD scheme is to build three TX/RX pa irs with three bond wire antennas (Fi gure A-9) on each side. This gives independence to the three I/O data paths trading off hardware sharing. The PA in each TX only needs to deal with one channel signal and the requirements on it are relaxed. Theoretically, a T/R switch allows one TX/RX working in their own time slots without interfering each other. But when three such transceivers are put t ogether on one chip, there could be a situation when one transceiver is receiv ing while the other two are transmitting, which presents the worst scenario for the one transcei ver trying to retrieve in formation from incoming signal, as shown in Figure A-10. The data rate is assumed to be 1 Gbps. Figure A-10. Adjacent bond wire antenna coupli ng causes serious adjacent channel interference problem and degrades received signal SNR. Signals (interference) c oupled from TX1 (Channel 1) and TX3 (Channel 3) to RX2 (Channel-2) are stronger than the desired Channel 2 signals. They c onstitute severe interferences and cause a multitude of problems in the system in Figure A-10. The coupling could be from the bond wire antenna, or through on-ch ip wiring or silicon substrat e. Among them, the bond wire antenna coupling probably will contribute the mo st while the other two seem to be more
173 manageable from the previous amplifier study in Chapter 3. A bond wire antenna coupling study shows that the coupling gain between two bond wire antenn as with a separation of 300 m are ~ dB around 60 GHz. Based on this, in-channel (Channel 2) SNR degradation caused by the coupling can be analyzed. From ADS simulations, a third order el liptical LPF at TX baseband is able to suppress the 2nd side lobe of the TX output BPSK si gnal to ~48 dB below the main lobe. Within the channel 2 in r eceiving mode, power coupled (PCW) from the TX1 and TX3 is estimated to be Equation A-3. PCW = TX1 Main Lobe Power + Coupling Gain -2nd Side Lobe Difference From Main Lobe After LPF Suppression-One Side L obe Factor + TX3 Coupling Factor = 3-8-48-3+3 = -53 dBm (A-3) This relatively large coupled interference power forces the desired received signal power to be at least -40 dBm to maintain the 13-dB SNR needed for the 10-16 BER. This consumes a large portion of the link margin (~20 dB) available at the best antenna pair gain. Interestingly, the inchannel coupled interference actually puts th e interconnect system output SNR (before ADC) limited by the interference instead of thermal noise similar to the situat ion in crowded cellular networks. In-channel SNR degradation can also be caused by reciproc al mixing between the Channel 2 LO carrier and the c oupled main lobes of Channel 1 and Channel 3 signals. Figure A11 illustrates such an issue. The coupled main lobe power in dB is estimated in Equation A-4. PCW = TX1 Main Lobe Power + Coup ling Gain + TX3 Coupling Factor = 3-8-48-3+3 = -53 dBm (A-4) When Channel 2 is used for reception, the required phase noise (PN) floor (dBc/Hz) of LO2 @ 2 GHz frequency offset can be estimated conservatively assuming the phase noise linearly reduces 6-dB from 2 to 4 GHz offset, as estimated in Equation A-5. PN
174 <-38-13-93-(-2) = 142 dBc/Hz (A-5) Figure A-11. Adjacent bond wire antenna coup ling causes serious reciprocal mixing problems and degrades received signal SNR It will be challenging to obtain such low phase noise fl oor with an on-chip mm-wave generator if not entirely impossible. What makes things even harder is the la rge out-of-channel coupled interference can satu rate the receiver front-end and generate large 2nd order and 3rd order inter-modulation products located in the desired ch annel. Unrealistic linea rity requirements will result from this. The coupled interference power of -2 dBm at RX2 input will require at least -2 dBm IP1dB for RX2 to avoid serious distortion and saturation. Two straightforward ways to improve the situation will be reducing unwanted bond wire coupling and using high Q tuning circuits for ea ch channel so that the out-of-channel coupled interference can be reduced. If the coupling gain drops from ~-8 to ~-25 dB, the receiver output SNR may improve ~17 dB assuming receiving dBm desired signal, the phase noise floor requirement will be relaxed by ~17 dB and IP1dB needed is reduced to -24 dBm. The undesirable bond wire coupling can be reduced by increasing the distance among neighboring bond wire antennas at th e expense of increased chip area. However, with the linear
175 dimension of modern VLSI chip reaching ~cm range, it is possibl e to separate the bond-wire antennas further than the 300 m assumed in previous discussion. The coupling can also be reduced by increasing the resonant frequenc y separations among bond wire antennas and Figure A-12. Alternative wireless FDMA interc onnect architecture increases hardware sharing and mitigates undesirable bond wire antenna coupli ng utilizing the bandpass filtering response of mm-wave TX/RX front-ends to attenuate unwan ted signals. To obtain maximal hardware sharing, an alternate TDD interconnect circuit system shown in Figure A-12 could be utilized. Only one PA and one LNA are needed for the th ree pairs of TX and RX and only one bond wire antenna is required on each side. This reduces the total number of antenna s and also allows the spacings among antennas to be increased. Thus, the mutual coupling among nearby bond wire antennas in the architecture of Figure A-9 can be reduced. Nevertheless, this system can only be used for the case where the thre e I/O data paths provided by the three wireless FDMA channels
176 are synchronized in terms of transmission and reception. The PA at mm-wave frequencies may have difficulty of supporting linear ity and output power for the thr ee channels at the same time. Three individual PA for each channel may have to be built and combined at the outputs. One key component for TDD scheme is the mm-w ave T/R switch. So far, it is assumed an ideal component in the above discussions. At this point, according to mm-wave T/R switch works done in our group using the TI CMOS pr ocess, ~60-70 GHz T/R switches have ~4 dB insertion loss (IL) and ~20 dB isolation. The IL is going to dissipate a la rge portion of PA output power and increase RX NF. This moderate perfor mance of T/R switch is a serious concern for the system in Figure A-12. A.9 Alternate Architecture with F requency Division Duplex Scheme Alternate system architecture based on FDD scheme and the associated radio circuit architecture for the wireless interconn ects is also possible (Figure A-13). The system does not need the non-ideal mm-wave T/R switch. The bidirectional communication in the system is supported using two different fr equency bands. On one side, 60 GHz band (low band) as reception and 80 GHz band (high band) for transmission. On the other side, the use of the frequency bands is reversed. For the radio circuit system architecture, PA and LNA can still be shared by three TXs/RXs. Unli ke the TDD scheme in Figure A-12, real-time bidirectional communication for three independent I/O data pa ths is possible with this architecture. Constructing bond wi re antenna that covers both th e low and high bands with more than 16-GHz bandwidth seems to be challenging. A high quality factor duplexer filter at mmwave is also a challenge. So instead of sharing one antenna, two bond wire antennas can be shared by the three TXs and th e three corresponding RXs respectively. If the PA cannot be shared by the three channels because of its perfor mance limits, then each TX will have to use its own PA and antenna. An antenna and an LNA in RX however can still be shared. All the TXs
177 and their antennas on one chip can be tuned to a different band from RXs. Hence the interference problem through bond wire antennas can be alleviated. Such an FDD scheme needs Figure A-13. Wireless FDMA inte rconnect system using an FDD sc heme utilizes two frequency bands for bidirectional communication, removes the T/R switch and reduces undesirable bond wire antenna coupling. more bond wire antennas and occupies two freque ncy bands instead of one compared to the TDD scheme shown in Figure A-12. Fo rtunately, adding bond wires shoul d be straightforward, while the frequency spectrum within the package (shielde d) is ample to allow the usage of two bands. The on-chip carrier generation sche me can be modified to provide carriers for two separate bands by adding multiple buffers, offset mixers and notch filters (Figure A-14) to the synthesizer architecture shown in Figure A-8 (B). Three LO signals at 78, 81 a nd 84 GHz (80 GHz band) will be generated. The 80 GHz band will have ~18 GHz difference from the original 60 GHz
178 band consisting of 60, 63 and 66 GHz carrier freq uencies. Inferred from the bond wire antenna coupling study that two antennas resonant 6 GHz apart could have a coupling gain of ~ dB, the coupling gain between two bond wire an tennas resonant at 18 GHz away (300 m apart) could be as low as ~-25 dB. Co mbining with the frequency tuning response of active circuits following the antenna, the isolation between the TX s and RXs could be higher than 40 dB. This high isolation should allevi ate all the issues facing TDD scheme s described in the last section. Table A-5 shows the system specification reflecting the adaptation of FDD scheme. Figure A-14. Multiple buffers, notch filters and offset mixers can be added to the 60-GHz band frequency synthesizer in Figure A-8 (B) to generate the 80-GHz band LO signals needed by the FDD scheme
179 Table A-5. Wireless inter-chip interconnect sy stem specification reflecting the adaptation of FDD scheme Features Specifications Data I/O path # 3 Multiple access FDMA FDMA channel # 3 Data rate/ channel (Gbps) 1 Modulation DPSK Duplex FDD RF BW/ channel (GHz) 2 Channel spacing (GHz) 3 Channel freq (GHz) 60,63,66,78,81,84 Bit error rate 10-16 SNR required (dB) 13.5 Total data rate (Gbps) 3 A.10 Summary In this chapter, a wireless interconnect system is proposed, targeting to alleviate the package and interconnect problems for future larg e scale computing circu its. First, the package and traditional wire interconnect challenges facing the digital circuits with increasingly higher speed and higher integration are presented. These challenges motivate multiple potential solutions and wireless interconnect is the one more compatible with mainstream CMOS technology and bringing lower cost To evaluate the feasibility of wireless interconnects in digital bulk CMOS, a mm-wave band FDMA inte rconnect system in a multi-chip module is proposed and studied. The system architecture, specification, and m odulation scheme are discussed, and the EM propagati on channel and radiation structures are introduced. Specifically, based on bond wire antennas measured characteristics, the interconnect li nk margin is estimated. Following this, direct-conversion radio circuit archite cture for the interc onnect system is
180 examined and the considerations for a multiple ou tput frequency synthesizer architecture design are discussed. The system duplex scheme in the context of the direct-con version architecture is re-examined. It is found that the system base d on TDD scheme could perform poorly due to nonideal performances of mm-wave T/R switc hes and undesirable coupling among bond wire antennas. Hence an alternate architecture using F DD scheme is proposed to mitigate these issues. The study in this chapter intends to identify general directio ns and potential issues of designing a wireless interconnect system. More detailed works in system simulation and circuit design are needed towards the re alization of such a system.
181 LIST OF REFERENCES  Institute of Electrical an d Electronics Engineers, IEEE WLAN Standard, Piscataway, NJ: IEEE, 2008.  Institute of Electrical an d Electronics Engineers, IEEE WPAN Standard, Piscataway, NJ: IEEE, 2008.  Institute of Electrical an d Electronics Engineers, IEEE WBAN Standard, Piscataway, NJ: IEEE, 2008.  G.E. Moore, Cramming more comp onents onto integrated circuits, Electronics, vol. 38, no. 8, pp. 114-117, Apr. 1965.  Semiconductor Industry A ssociation, ITRS, Int. Technol. Roadmap For Semiconductors, 2007 Edition, System Drivers, San Jose, CA: SIA, 2009.  K. X. Zhang, The challenges and opport unities in GHz microprocessor design on 0.13 m and beyond technologies, in Proc. of 6th Int. Conf. on Solid-State and Integrated-Circuit Tech., Oct. 2001, pp. 1102-1106.  J.-H. Park, Gb MLC (multi-level cell) NAND flash memory using 63 nm process technology, in IEEE IEDM Tech. Dig., Dec. 2004, pp 873876.  V. Balan, A 4.8.4-Gb/s Serial Link for Backplane Ap plications Using Decision Feedback Equalization, IEEE J. Solid-State Circuits, vol. 40, no. 9, pp 1957-1967, Sep. 2005.  R. B. Staszewski, A 24mm2 Quad-Band Single-Chip GSM Radio with Transmitter Calibration in 90nm Digital CMOS, in IEEE ISSCC Tech. Dig., Feb. 2008, pp 208300.  D. Seippel, M. Hammes, A. Hanke, and J. Kissing, GSM/GPR S Single-Chip in 130nm CMOS: Challenges on RF for SoC Integration, in IEEE RFIC Symp. Tech. Dig., Jun., 2006, pp. 4-180.  M.-C. F. Chang, Impact of CMOS Scali ng on RF/Mixed-Signal Circuit Designs, in Proc. of 8th Int. Conf. on Solid-State and Integrated Circuit Tech., May 2006, pp. 200-204.  R. B. Staszewski, K. M uhammad and D. Leipold Digital RF Processor Techniques for Single-Chip Radios (Invited), in IEEE CICC Tech. Dig., Sep. 2006, pp 789-796.  J. Mitola, The Software Radio Architecture, IEEE Comm. Magazine, vol. 33, no. 5, pp. 26-38, May 1995.
182  M. Feng, Device Technologies for RF Fr ont-End Circuits in Next-Generation Wireless Communications, Proc of IEEE, vol. 92, no. 2, pp 354-375, Feb. 2004.  H. S. Bennett, Device and Technology Evolution for Si-Based RF Integrated Circuits, IEEE Trans. Electron Dev., vol. 52, no. 7, pp. 1009-1019, Jul. 2005  S. K. Yong and C.-C. Chong, An Overvi ew of Multigigabit Wireless through Millimeter Wave: Technology: Poten tials and Technica l Challenges, EURASIP J. Wireless Comm. Networking, vol. 2007, pp 101-121, Dec. 2007.  T. Cutler, Unlicensed Wireless Data Communications, Compliance Engineering Magazine, pp. 12-16, Jan. 2002.  D. Cabric, M. S.W. Chen, D. A. Sobel, S. Wang, J. Yang, and R. W. Brodersen, Novel Radio Architectures for UWB, 60 GHz, and Cognitive Wireless Systems, EURASIP J. Wireless Comm. Networking, vol. 2006, pp. 21-41, Dec. 2007.  System Support Solutions Inc., BridgeWave 60-GHz Products Brochure, Orono, MN: SSS, 2008.  N. Guo, R. C. Qiu, S. S. Mo, and K. Takahashi, GHz Millimeter-Wave Radio: Principle, Technology, and New Results, EURASIP J. Wireless Comm. Networking, vol. 2007, pp. 51-81, Dec. 2007.  P. Smulders, Exploiting the 60 GHz Band for Local Wireless Multimedia Access: Prospects and Future Directions, IEEE Comm. Magazine, vol. 33, no. 5, pp. 140147, Jan, 2002.  J. W, Meyer, Millimeter Wave Research : Past, Present and Future, Lincoln Laboratory Technical Report 389, Boston, MA: MIT, May 1965.  A. Hajimiri, mm-Wave Silicon ICs: Challenges and Opportunities, in IEEE CICC Tech. Dig., Sep. 2007, pp. 741-747.  R. A. Pucel, Design Considerations for Monolithic Microwave Circuits, IEEE Trans. Microw. Theory Tech., vol. 29, no. 6, pp 513-534, Jun. 1981.  D. Yu, Ultra High-Speed 0.25m Emitter InP-InGaAs SHBTs with fmax of 687 GHz, in IEEE IEDM Tech. Dig., Dec.,2004, pp. 557-560.  K. Ohata, .25Gbps Wireless Gigabit Ethernet Link at 60 GHz-Band, in IEEE IMS Tech. Dig., Jun. 2003, pp. 373-376.  H. Zirath, C. Fager, M. Ga rcia, P. Sakalas, L. Lande n, and A. Alping, Analog MMICs for Millimeter-Wave Applica tions Based on a Commercial 0.14m pHEMT Technology, IEEE Trans. Microw. Theory Tech., vol. 49, no. 11, pp. 1312-1322, Nov. 2001
183  S. E. Gunnarsson, Highly Integrated 60 GHz Transmitter and Receiver MMICs in a GaAs pHEMT Technology, IEEE J. Solid-State Circuits, vol. 40, no. 11, pp 21742186, Nov. 2005.  E. Laskin, -GHz Transceiver with On -Chip Antennas in SiGe Technology, in IEEE RFIC Symp. Tech. Dig., Jun. 2008, pp. 637-640.  S. T. Nicolson, A 1.2V, 140GHz receiver with on-die antenna in 65nm CMOS, in IEEE RFIC Symp. Tech. Dig., Jun. 2008, pp 229-232.  J. S. Kilby, Miniaturized Electronic Circuits, United States Patent 3138743, issued on Jun. 23, 1964.  J. S. Kilby, Method of Making Mi niaturized Electronic Circuits, United States Patent 3261081, issued on Jul. 19, 1966.  R. N. Noyce, "Microelectronics," Scientific American, vol. 237, no. 3, pp. 62-69, Sep. 1977.  E. Laskin and S. Voinigescu, Towards 100+ GHz Electronics, in CMOSET Tech. Dig., Sep., 2007, pp 150-153.  B. Gaucher, Progress in SiGe Technology Toward Fully Integrated mmWave Ics, in IEEE ISTDM Tech. Dig., Oct. 2006, pp. 1-2.  B. Floyd, S. Reynolds, U. Pf eiffer, T. Zwick, T. Beuke ma, and B. Gaucher, SiGe Bipolar Transceiver Circu its Operating at 60GHz, IEEE J. Solid-State Circuits, vol. 40, no. 1, pp. 156-167, Jan. 2005.  A Babakhani, X Guan, A Komijani, A Nata rajan, and A. Hajimiri, A 77-GHz Phased-Array Transceiver with On-Chip antennas in silicon: Receiver and Antennas, IEEE J. Solid-State Circuits, vol. 41, no. 12, pp. 2795-2806, Dec. 2006.  S. Reynolds, Second Generation 60-GHz Tr ansceiver Chipset Supporting Multiple Modulations at Gb/s data rates (Invited), in IEEE BCTM Tech. Dig., Sep. 2007, pp. 192-197.  J. Pekarik, RF CMOS Technology from 0.25 m to 65 nm: The State of the Art, in IEEE CICC Tech. Dig., Sep. 2004, pp. 217-224.  C. H. Doan, Millimeter-wave CMOS design, IEEE J. Solid-State Circuits, vol. 40, no. 3, pp. 144 -155, Jun. 2005.  T. Yao, -GHz PA and LNA in 90-nm RF-CMOS, in IEEE RFIC Symp. Tech. Dig., Jun., 2006, pp. 100-103.
184  S. Emami, C.H. Doan, A.M. Niknejad, R. W. Brodersen, A 60-GHz DownConverting CMOS Singl e-gate Mixer, in RFIC Symp. Tech. Dig., June, 2005, pp. 163-166.  B. Razavi, CMOS Transceivers for the 60-GHz Band, in IEEE RFIC Symp. Tech. Dig., Jun., 2006, pp. 106-109.  Semiconductor Industry Association, Int. Technol. Roadmap For Semiconductors, 2005 Edition, Radio Frequency And Analog/Mixed-Signal Technologies For Wireless Communications, San Jose, CA: SIA, 2005.  S. E. Thompson, A 90-nm Logic T echnology Featuring Strained-Silicon, IEEE Trans. Electron Dev., vol.51, no. 3, pp.1790-1797, May 2004.  T. O. Dickson, The Invariance of Characteristic Current Densities in Nanoscale MOSFETs and Its Impact on Algorithmic De sign Methodologies and Design Porting of Si(Ge) (Bi)CMOS HighSpeed Building Blocks, IEEE J. Solid-State Circuits, vol. 41, no. 8, pp. 1006-1016, Aug. 2006.  R. Dennard, Design of ion-implante d MOSFETs with very small physical dimensions, IEEE J. Solid-State Circuits, vol. 9, no. 5, pp. 256-268, Oct. 1974.  Y. Taur and T. H. Ning, Modern VLSI Devices, New York, NY: Cambridge University Press, 1998.  L. M. Franca-Neto, Fully Integrated CM OS Radios from RF to Millimeter Wave Frequencies, Intel Tech. J., vol 08, issue 03, pp. 1-10, Aug. 2004.  BSIM Group, BSIM 4.3.0 MOSFET Model-Users Manual, Berkeley, CA: University of California, 2003.  S. F. Tin, A. A. Osman, K. Mayaram, a nd C. Hu, A Simple Subcircuit Extension of the BSIM3v3 Model for CMOS RF Design, IEEE J. Solid-State Circuits, vol. 35, no. 4, pp.612-624, Apr. 2000.  C. H. Doan, S. Emami, A. M. Niknejad, and R. W. Brodersen, Millimeter-Wave CMOS Device Modeling And Simulation, in IEEE ISCAS Tech. Dig., May. 2004, pp. 524-527.  C. H. Doan, S. Emami, A. M. Niknejad, and R. W. Brodersen, Design of CMOS for 60 GHz applications, in IEEE ISSCC Tech. Dig., Feb. 2004, pp. 440-538.  M. W. Pospieszalski, On the measuremen t of noise parameters of microwave twoports, in IEEE IMS Tech. Dig., Apr. 1986, pp.456-458.  C. H. Doan, S. Emami, A. M. Niknejad, and R. W. Brodersen, Millimeter-Wave CMOS Design, IEEE J. Solid-State Circuits, vol. 40, no. 1, pp. 144-155, Jan. 2005.
185  C. Enz, and Y. Cheng, MOS Tran sistor Modeling for RF IC Design, IEEE Trans. Solid-State Circuits, vol.35, no.2, pp 186-123, Feb. 2000.  S.-M. Yim, COMPONENTS AND CIRCUIT ANALYSIS TECHNIQUES FOR A MULTI-BAND WIRELESS RECEIVER, PhD Thesis, UNIVERSITY OF FLORIDA, Gainesville, FL, 2003.  T. H. Lee, The Design of CMOS Radio-Fr equency Integrated Circuits, New York, NY: Cambridge University Press, 1998.  Ansoft LLC., HFSS User Guide, Version 10.0, Pittsburgh, PA: Ansoft LLC., 2005.  C-M. Hung, Y.-C. Ho, I.-C. Wu, and K. K. O, High-Q capacitors implemented in a CMOS low-power wireless applications, IEEE Trans. Microw. Theory Tech., vol. 46, no. 5, pp. 505, May 1998.  N. Zhang and K. K. O, 94 GHz Voltage Controlled Oscillator with 5.8 % Tuning Range in Bulk CMOS, IEEE Microw. Wireless Compon. Lett., vol. 15, no. 8, pp.493-495, Aug. 2008.  C. P. Yue and S. S. Wong, On-Chip Spir al Inductors with Pa tterned Ground Shields for Si-Based RF ICs, IEEE J. Solid-State Circuits, vol. 33, no. 5, pp. 743-752, May 1998.  FastHenry Group, FastHenry Users Guide, Version 3.0, Boston, MA: MIT, 1996.  K. K. O, Wireless Communication Using Integrated Antennas, in Proc. Int. Interconnect Technol.. Conf., Jun. 2003, pp.111-113.  B. A. Floyd, A CMOS WIRELE SS INTERCONNECT SYSTEM FOR MULTIGIGAHERTZ CLOCK DISTRIBUTION, Ph.D. Thesis, UNIVERSITY OF FLORIDA, Gainesville, FL, May 2001.  U. R. Pfeiffer, Low-loss Contact Pad with Tuned Impedance for Operation at Millimeter Wave Frequencies, in SPI Tech. Dig., May 2005, pp 61-64.  G. Gonzalez, Microwave Transistor Amplifier: Analysis and Design, 2nd edition, New York, NY, Pearson Print on Demand, 1996.  D. K. Shaeffer and T. H. Lee, A 1.5-V, 1.5-GHz CMOS low noise amplifier, IEEE J. Solid-State Circuits, vol. 32, no.5, pp. 745-759, May. 1997.  Y.-C. Ho, M. Biyani, J. Colvin, C. Smit hhisler, and K. K. O, A 3-V Low Noise Amplifier Implemented Using a 0.8m CMOS Process with Three Metal Layers for 900 MHz Operation, IEE Electronics Lett., vol. 32, no. 13, pp. 1191-1193, Jun. 1996.
186  C. H. Doan, S. Emami, A. M. Niknejad, and R. W. Brodersen, Design of CMOS for 60 GHz applications, in IEEE ISSCC Tech. Dig., Feb. 2004, pp. 440.  C.-M. Lo, C.-S. Lin, and H. Wang, A Miniature V-band 3-St age Cascode LNA in 0.13m CMOS, in IEEE ISSCC Tech. Dig., Feb. 2006, pp 1254-1255.  S. Emami, C.H. Doan, A.M. Niknejad and R.W. Brodersen, A highly integrated 60 GHz CMOS front-end receiver, in IEEE ISSCC Tech. Dig., Feb. 2007, pp. 190.  T. Mitomo, R. Fujimoto, N. Ono, R. Tach ibana, H. Hoshino, Y. Yoshihara, Y. Tsutsumi, and I. Seto, A 60-GHz CMOS Receiver Front-End With Frequency Synthesizer, IEEE J. Solid-State Circuits, vol. 43, no. 4, pp. 1030, April 2008.  K.-H. Chen, C. Lee and S.-I. Liu, A Dual-band 61.4~63GHz/75.5~77.5GHz CMOS Receiver in a 90nm Technology, in VLSI Symp. on Circuits Tech. Dig., June 2008, pp. 160.  M. Khanpour, K. W. Tang, P. Garcia, a nd S. P. Voinigescu, A Wideband W-Band Receiver Front-End in 65-nm CMOS, IEEE J. Solid-State Circuits, vol. 43, no. 8, pp. 1717, Aug. 2008.  E. Laskin, M. Khanpour, R. Aroca, K.W. Ta ng, P. Garcia, and S. P. Voinigescu, A 95GHz receiver with fundamental-frequency VCO and static frequency divider in 65nm digital CMOS, in IEEE ISSCC Tech. Dig, Feb. 2008, pp. 180.  N. Zhang, C.-M. Hung, and K. K. O, 80-GHz Tuned Amplifier in Bulk CMOS, IEEE Microw. Wireless Compon Lett., vol.18, no.2, pp. 121-123, Feb. 2008.  B. A. Floyd, J. Mehta, C. Gamero, and K. K. O, A 900-MHz, 0.8m CMOS low noise amplifier with 1.2-dB noise figure, in IEEE CICC Tech. Dig., May, 1999, pp. 661-664.  Jri Lee, High-Speed Circuit Designs for Transmitters in Broadband Data Links, IEEE J. Solid-State Circuits, vol. 41, no. 5, pp 1004-1015, May 2006.  B. Razavi, RF Microelectronics, Pearson Education Asia Limited, 2003, pp. 212.  F. Y. Yuan, Electromagnetic Modeli ng and Signal Integrity Simulation of Power/Ground Networks in High Speed Di gital Packages and Printed Circuit Boards, in DAC Tech. Dig. Jun. 1998, pp. 300.  T. Blalack, J. Lau, F. J. R. Clement, a nd B. A. Wooley, Experimental results and modeling of noise coupling in a lightly doped substrate, in IEEE IEDM Tech. Dig., Dec. 1996, pp. 623.  A. Samavedam, A. Sadate, K. Mayaram, and T. S. Fiez, A scalable substrate noise coupling model for design of mixed-signal ICs, IEEE J. Solid-State Circuits, vol. 35, no. 6, pp. 895, Jun. 2000.
187  T. Blalack, Y. Leclercq, and C.P. Yue, On-Chip RF Isolation Techniques, in IEEE BCTM Tech. Dig., Oct. 2002, pp. 205.  Agilent Inc., Agilent PNA E8361A Manual, Santa Clara, CA: Agilent Inc., 2008.  Z. Shen, Advanced Microwave/Millimeter-Wave Imaging Technology, J. Plasma Fusion Research, vol. 2, pp. 1019-1023, May 2007.  L. M. Franca-Neto, R. Bi shop, and B. Bloechel, GHz and 100 GHz VCOs in 90 nm Using Optimum Pumping Method, in IEEE ISSCC Tech. Dig., Feb.2004, pp. 444-445.  C. Cao and K.K.O, A 90-GHz voltage-controlled oscillat or with a 2.2-GHz tuning range in 130-nm CMOS technology, in Symp. VLSI Circuits Tech. Dig., Jun. 2005, pp. 242.  C. Cao and K.K.O, Millimeter-Wave Voltage-Controlled Oscillators in 0.13 m CMOS Technology, IEEE J. Solid-State Circuits, vol. 41, no. 6, pp 1297-1304, Jun. 2006.  C. Cao and K.K.O, A 140-GHz Fundamenta l Mode Voltage-Controlled Oscillator in 90-nm CMOS Technology, IEEE Microw. Wireless Compon. Lett., vol. 16, no. 10, pp 555-557, Oct. 2006.  Z.-M. Tsai; C.-S. Lin; Huang, C.F. Chern, and H. Wang, A Fundamental 90-GHz CMOS VCO Using New Ring-Coupled Quad, IEEE Microw. Wireless Compon. Lett., vol. 17, no. 3, pp 226-228, Mar. 2007,  P.-C. Huang, M.-D. Tsai, H. Wang, C.-H. Chen, and C.-S. Chang, A 114 GHz VCO in 0.13 m CMOS technology, in IEEE ISSCC Tech. Dig., Feb.2005, pp. 404-405.  P. C. Huang, R.-C. Liu, H.-Y. Chang, C.-S Lin, M.-F. Lei, H. Wang, C.-Y. Su, and C.-L. Chang, A 131 GHz push-push VC O in 90-nm CMOS technology, in IEEE RFIC Symp. Tech. Dig., Jun. 2005, pp 613.  C. Cao, E. Seok and K.K.O, 192 GHz push-push VCO in 0.13 m CMOS, IEE Electronics Lett., vol. 42, no.4, pp 208-210, Feb 2006.  K. Ishibashi, M. Motoyoshi, N. Kobayashi, and M. Fujishima GHz CMOS Voltage-Controlled Oscillator with 7% Frequency Tuning Range, in Symp. VLSI Circuits Tech. Dig., Jun. 2007, pp. 176.  C, Lam, Phase-locked loop with self-selecting multi-band VCO, United States Patent 6806786, issued on Oct. 19, 2004.  J. J. Hung, T. M. Hancock, and G.M. Rebeiz, A 77 GHz SiGe Sub-Harmonic Balanced Mixer, IEEE J. Solid-State Circuits, vol. 49, no.11, pp. 2167-2173, Nov. 2005.
188  R. M. Kodkani and L.E. Larson, An Integrated 50-GHz SiGe Sub-Harmonic Mixer/Downconverter with a Quadrature Ring VCO, in IEEE Topical Meeting Si. Mono. Integr. Circuits in RF Systems Dig., Jan. 2007, pp 223-226.  J. -H. Tsai and T. -W. Huang, 65-GHz CMOS Broadband Modulator and Demodulator with Sub-harmonic Pu mping for MMW Wireless Gigabit Applications, IEEE Trans. Microw. Theory Tech., vol. 55, no. 10, pp.2075-2085, Oct, 2007.  N. Zhang, H. Xu, H.-T. Wu, and K. K. O, W-band Active Down-Conversion Mixer in Bulk CMOS, IEEE Microw. Wireless Compon Lett., vol. 19, no. 2, pp. 98-100, Feb. 2009.  Maxim Integrated Products Inc., Thr ee Methods of Noise Measurement, Application Note 2875, Dallas, TX, Nov. 21, 2003.  B. M. Motlagh S. E. Gunnarsson, M. Ferndanl, and H. Zirath, Fully Integrated 60GHz Single-ended Resistive Mixe r in 90-nm CMOS Technology, IEEE Microw. Wireless Compon. Lett., vol.16, no.1, pp 25-27, Jan. 2006.  C.-S Lin, P.-S Wu, H.-Y Chang, and H. Wang, A 9-50-GHz Gilbert-cell Downconversion Mixer in 0.13m CMOS Technology, IEEE Microw Wireless Compon. Lett., vol.16, no.5, pp 293-295, May. 2006.  M. Varonen, M. Karkkainen, and K. Halone n, V-band Balanced Resistive Mixer in 65-nm CMOS, in Europe. Solid-State Circ. Conf. Tech. Dig., Sep, 2007, pp 360-363.  Institute of Electrical and Electronics Engineers, Virtual Museum Radar Introduction, Piscataway, NJ: IEEE, 2008.  H. Rohling, Research Activiti es In Automotive Radar, in Symp. MSMW Tech. Dig., Jun., 2001, pp. 155-158.  K.M. Strohm, H.-L. Bloecher, R. Schneider, and J. Wenger, Development of future short range radar technology, in Proc. Europe. Radar Conf., Oct. 2005, pp. 165-168.  Directorate General Info rmation Society, A 2nd Generation Forward Looking Radar Sensor News for Adaptive Cruise Contro l, The Directorate C Lett. Mini. Embed. Systems App., vol.2, no.4, pp 2, May 2004.  W. L. Stutzman and G. A. Thiele, Antenna Theory and Design, 2nd Edition, New York, NY: Wiley, 1998.  X. Li, T. Brogan, M. Esposito, B. Myers a nd K. K. O, A Comparison of CMOS and SiGe LNAs and Mixers for Wi reless LAN Application, in IEEE CICC Tech. Dig., May, 2001, pp. 531-534.
189  Y. Deguchi, K. Sakakibara N. Kikuma, and H. Hirayama, Millimeter-wave microstrip-to-waveguide transition operating over broad frequency bandwidth, in IEEE IMS Tech. Dig., Jun. 2005, pp. 2107-2110.  C. Cao, Y. Ding, and K. K. O, A 50-GHz Phase-Locked Loop in 0.13m CMOS, IEEE J. Solid-State Circuits, vol. 42, no. 8, pp 1649-1656, Aug. 2007.  C. Cao and K. K. O, A Power Efficien t 26-GHz 32:1 Static Frequency Divider in 130-nm Bulk CMOS, IEEE Microw Wireless Compon. Lett., vol. 15, no. 11, pp 233235, Nov. 2005.  R. Mohanavelu and P. Heydari, A N ovel 40-GHz Flip-Fl op-Based Frequency Divider in 0.18 mm CMOS, in Proc. of ESSCIRC, Sep. 2005, pp. 185-188.  Yanping Ding, CMOS SYNTHESIZERS FOR 24-GHZ RADIO, Ph.D. Thesis, UNIVERSITY OF FLORIDA, Gainesville, FL, 2007.  C. Cao, A 24-GHz FULLY-INTEGRATED CMOS TRANSMITTER WITH ONCHIP ANTENNA, Ph.D. Thesis, UNIVERSITY OF FLORIDA, Gainesville, FL, 2006.  C. Lee and S.I. Liu, A 58-to-60.4GHz Frequency Synthesizer in 90nm CMOS, in IEEE ISSCC Tech. Dig., Feb. 2007, pp. 196-197.  H. Hoshino et al., A 60-GHz Phase-Locked Loop with Inductor-le ss Prescaler in 90nm CMOS, in Proc. of ESSCIRC, pp. 472-475, Sep. 2007.  J. Lee, M. Liu and H. Wang, A 75GHz Phase-Locked Loop in 90-nm CMOS Technology, IEEE J. Solid-State Circuits, vol. 43, no. 6, pp. 1414-1426, June 2008.  K.-H. Tsai and Shen-Iuan Liu, A 43.7mW 96GHz PLL in 65nm CMOS,, in IEEE ISSCC Tech. Dig., Feb. 2009, pp. 276-277.  K. Scheir, G. Vandersteen, Y. Rolain, a nd P. Wambacq, A 57-to-66GHz Quadrature PLL in 45nm Digital CMOS, in IEEE ISSCC Tech. Dig., Feb. 2009, pp. 494-495.  Vipul Jain, Fred Tzeng, Lei Zhou and Pa yam Heydari, A Singl e-Chip Dual-Band 22-to-29GHz/77-to-81GHz BiCMOS Tran sceiver for Automotive Radars, in IEEE ISSCC Tech. Dig., Feb. 2009, pp. 308-309.  Yoichi Kawano et al, A 77GHz Transceiver in 90nm CMOS, in IEEE ISSCC Tech. Dig., Feb. 2009, pp. 310-311.  K. K. O, Wireless Inter-chip Interconnect System, Proposal to NSF, UNIVERSITY OF FLORIDA, Gainesville, FL, 2004.  H. K. Kwon, SIP solution for high-end multimedia cellular phone, in IEEE IMAPS Tech. Dig., Sep. 2003, pp.165-169.
190  T. Oka, System-In-a-Package Stacked Flash Memory and Logic LSI, in Semicon Japan, 2000.  N. Miura, An 11Gb/s Inductive-Coupling Link with Burst Transmission, in IEEE ISSCC Tech. Dig., Feb. 2008, pp 298-614.  A. V. Krishnamoorthy, Optical interconne cts for commodity silicon technologies, in IEEE Int. Conf. Innovative Sy stems In Silicon Tech. Dig., Oct 1997, pp 201-202.  J. Ko, J. Kim, Z. Xu, Q. Gu, C. Chien, and M.F. Chang, An RF/Baseband FDMAInterconnect Transceiver for Reconfigur able Multiple Access Chip-to-Chip Communication, in IEEE ISSCC Tech. Dig., Feb. 2005, pp 208-211.  H.-T. Wu, M. Tekle, C.S. Nallani, N. Zhang and K. K. O, Bond Wire Antennas for Operations Near 60 GHz, IEEE Trans. Microwave Theo. Tech., Submitted.  C. Sandner and A. Wiesbauer, A 3 GHz to 7 GHz fast -hopping frequency synthesizer for UWB, in Proc. Int. U ltra Wideband Syst. Workshop, 2004, pp. 405 409.  D. Huang, A 60GHz CM OS VCO Using On-Chip Resonator with Embedded Artificial Dielectric for Size, Loss and Noise Reduction, in IEEE ISSCC Tech. Dig., Feb. 2006, pp 314-315.
BIOGRAPHICAL SKETCH Ning Zhang received his B.S. degree from th e Dept. of Electron ics and Communication Engineering, South China University of Tec hnology, Guangzhou, China in 2001 and his M.S. degree from the Dept. of Electrical and Comput er Engineering, University of WisconsinMadison in 2004 with research in THz micro-vacuum devices de sign. He received his Ph.D. degree from the Dept. of Electrica l and Computer Engineering, University of Florida in August 2009, focusing on CMOS millimeter-wave/RF circui ts and systems and their applications. He worked on inductor extraction, ci rcuit simulators benchmarking, and power amplifier package modeling as an intern in Wire less Branch of Conexant Inc. Melbourne, Florida in 2007.