CMOS Millimeter Wave Receivers

ARENBERG DOCTORAL SCHOOL Faculty of Engineering Science

CMOS Millimeter Wave Receivers

Maarten TYTGAT Promotor: Prof. dr. ir. ing. P. Reynaert

Dissertation presented in partial fulfillment of the requirements for the degree of Doctor in Engineering

Co-promotor: Prof. dr. ir. M. Steyaert

May 2014

CMOS Millimeter Wave Receivers

Maarten TYTGAT

Examination Committee: Prof. dr. ir. H. Hens, chair Prof. dr. ir. ing. P. Reynaert, promotor Prof. dr. ir. M. Steyaert, co-promotor Prof. dr. ir. W. Dehaene Prof. dr. ir. R. Lauwereins Prof. dr. ir. P. Leroux ir. W. De Raedt (IMEC) M. Keaveney, BE, MEng (Analog Devices)

Dissertation presented in partial fulfillment of the requirements for the degree of Doctor in Engineering

May 2014

© KU Leuven – Faculty of Engineering Science Kasteelpark Arenberg 10, 3001 Heverlee (Belgium) Alle rechten voorbehouden. Niets uit deze uitgave mag worden vermenigvuldigd en/of openbaar gemaakt worden door middel van druk, fotocopie, microfilm, elektronisch of op welke andere wijze ook zonder voorafgaande schriftelijke toestemming van de uitgever. All rights reserved. No part of the publication may be reproduced in any form by print, photoprint, microfilm or any other means without written permission from the publisher. D/2014/7515/67 ISBN 978-94-6018-843-5

Voorwoord

Na een goeie honderd-en-een-klets pagina’s in het Engels te hebben getypt, is het bijna een verademing om nog enkele pagina’s in het Nederlands te mogen typen. Het is tevens een welgekome afwisseling om niet om de vijf woorden een LATEX tag te moeten gebruiken (beste backslash, we hebben een mooie tijd gehad, maar nu is het even genoeg geweest). In dit voorwoord wil ik graag mijn best doen om bij het bedanken van de vele mensen die mijn doctoraat mee tot een goed einde (hout vasthouden) hebben gebracht, niemand te vergeten. Een chronologische volgorde kan daar misschien bij helpen. . . Het was tijdens mijn thesisjaar dat ik in contact kwam met professor Patrick Reynaert. Die thesis ging al een beetje over millimeter waves, en ik vond dat wel chique, zo een beetje op de toenmalige final frontier van de elektronica werken (frequentiegewijs). Maar naar het schijnt heb ik toch ooit tegen mijn thesisbegeleider Brecht Machiels gezegd dat ik absoluut niet wou doctoreren. Zo een hele dag in een kelder naar een PC-scherm zitten staren, nee bedankt. Ik zou wel wachten tot een of ander hip bedrijf mij zou opbellen om er te komen werken. Helaas was de crisis toen net begonnen en bleven zulke telefoontjes uit. Intussen deed ik lustig voort aan mijn thesis en besefte ik dat ik dat eigenlijk wel leuk vond, zo op het gemak een technisch probleem tot op het bot (dacht ik) onderzoeken en met een oplossing (dacht ik) op de proppen komen. Bovendien liep ik toen regelmatig eens door de MICAS-kelder en beviel die sfeer mij wel. Toen Patrick me op een gegeven moment vroeg om een doctoraat te starten, heb ik ook niet heel erg lang getwijfeld (denk ik). Om dus op dat bedanken terug te komen: Patrick is de eerste in het rijtje, niet alleen omwille van de chronologische volgorde, maar ook, en vooral, omwille van de tijd die hij altijd voor me uittrok om mijn onderzoek te bespreken. De weekly meetings zijn dan misschien nooit een succes geworden, maar ik denk dat als we het uitmiddelen, we gemakkelijk aan een meeting per week komen. Naast de weekly meetings was er ook elk jaar een Summer Group Meeting voor alle doctoraatsstudenten van Patrick, met barbecue en de nodige hoeveelheid bier, die naast heel gezellig, ook uitermate geschikt was om de collega’s met een andere promotor jaloers te maken (Veronique, ook bedankt!). Hoewel Patrick tegenwoordig de reputatie heeft dat alles wat hij aanraakt in het labo, stopt met werken, is dat niet altijd zo geweest. Bij de eerste meting van mijn eerste chip, is hij tot ’s avonds laat gebleven om te helpen met de meetopstelling, hoopvol te wachten op een piekje op de spectrum analyzer en te delen in de vreugde. i

ii

Michiel Steyaert heeft mij als co-promotor ook erg geholpen. Alhoewel ik van meetings met hem vaak terugkwam met meer vragen dan ik op voorhand had, bleken die extra vragen niet alleen zichzelf maar ook de originele vragen vaak te beantwoorden. De paper die hij tijdens een meeting ooit volgepend heeft met cryptische boodschappen, zal ik met zorg bewaren (behalve als ik er ooit veel geld mee kan verdienen op een veiling). Een jury is er om credibiliteit te geven aan een doctoraat. Zonder hen zou dit boekje niets waard zijn. Graag bedank ik de voorzitter van de jury, Hugo Hens, om de zaken in goede banen te leiden, Wim Dehaene, Rudy Lauwereins, Paul Leroux, Walter De Raedt en Mike Keaveney voor het grondig nalezen van de tekst, het stellen van terechte vragen en aanleveren van nuttige suggesties. Thank you Mike, for coming all the way to Leuven to attend the preliminary defense and for the very useful comments. Zoals ik eerder al zei, was het misschien wel de sfeer op MICAS die mij overhaald had om er aan een doctoraat te beginnen. Die sfeer wordt wel eens uniek genoemd. Ik vond dat altijd nogal voorbarig, omdat ik nog nooit ergens anders had gewerkt, maar hoe meer ik met andere doctoraatstudenten (niet-MICASsers dus) praat, hoe meer ik besef wat een chance dat wij hier hebben. Ik denk dat alle MICAS-proffen hiervoor verantwoordelijk zijn en bij deze wil ik dus ook hen bedanken voor het toelaten en zelfs aanmoedigen van —zacht uitgedrukt— niet-wetenschappelijke activiteiten zoals recepties, kanonschieten, kubben, recepties, quizzen, koffiepauzes en recepties. Wim wil ik graag nog extra bedanken voor de fijne samenwerking bij Elektrische Netwerken, Solders of Fortune en diverse muzikale projecten. Uiteraard zijn het ook de collega’s zelf die deze sfeer in stand houden. Daarom wil ik hen graag in quasi-chronologische volgorde oplijsten, samen met hun belangrijkste verdiensten: Stefan Cosemans (Matlab Java stuff), Frederik Ceyssens (algemene Micasheld, altijd overal zijn), Mike Wens (nieuwelingen op hun plaats zetten), Koen Cornelissens, Jef Thoné, Filip Tavernier (dingen duidelijk uitleggen), Zheng Li (singing with the MICAS band), Lianming Li (best public defense ever), Christophe De Roover, Jorg Daniels, Tom Van Breussegem (de mail naar de rector, uw groot bakkes), Hagen Mariën (legendarische fishing trip!), Philippe Jourand, Brecht Machiels (voor de thesis en het droge cynisme), Anselme Vignon, Bo Liu (best dreams and made-up stories), Cedric Walravens, Elie Maricau (beste receptie), Pieter De Wit, Pieter Nuyts (altijd in voor een goeie discussie), Hans Danneels (het trompetincident, de vistrip), Dimitri De Jonghe (uw Rummicub - Jenga cross-over spel), Fridolin Michel (still using your USB-stick), Ying Cao, Noël Deferm (zie later), Valentijn De Smedt (MICAS band, Solders of Fortune, nog heel veel..), Xinpeng Xing, Bram Roseleer (de grotten en het ijs), Nele Reynders (the original Kubb), Jens Verbeeck (was toch nog laachen in Aachen), Tom Redant (the music, het Limburg-gevoel) Brecht François, Ercan Kaymaksut, Hans De Clercq (whiskey, whisky), Hans Meyvaert (whisky, appelsientruuks, algemene hilariteit), Wouter Volkaerts (zie later), Tiannan Guan, Adi Xhakoni, Maarten Strackx (trendsetter met speelgoed in ESAMAT), Dixian Zhao (keeping an eye on Niels), Shailesh Kulkarni (best thesis student, best Indian food!), Henk Motte, Piet Callemeyn

iii

(trouwe feestvierder, goochel- en saxheld), Nico De Clercq (beste cakekasbeheerder), Grim Keulemans, Paramartha Indirayanti, Iman Khajenasiri, Jeroen Lecoutere (politiek advies, de zalm met aardappelen), Pieter Gijsenbergh (sputteren), Wouter Steyaert (het laten staan van uw baard), Jelle Van Rethy (ten gepaste tijde op de grond tuffen), Hans Reyserhove (Kempenzoon eerste klas), Aki Sarafianos (beste verjaardagstaarten), Marco Vigilante (best drunk bass player), Luigi Brancato (helping with that nasty splinter), Patrick Pelgrims, Niels Van Thienen (goeie tijden met die extruder!), Juan Carlos Pena (beste Nederlands), Anthony Coyette (tweede beste Nederlands, second best dance moves), Jolien De Meester (de positiviteit en het zingen), Ha Le-Thai, Jorge Marin (best drunk solo artist), Tuba Ayhan, Burak Baran, Peng Zhu, Marco Simicic (best dance moves), Komail Badami, Nicolas Butzen, Florian De Roose, Baris Esen, Monique Ingels, Ibrahim Kazi, Hakim Kobbi, Steven Lauwereins, Fernando Rosas, Bert Moons (hilarisch tuinkersmoment), Robin Theunis (beste Nieuweling), Yang Zhang, Buddhadev Paul Chaudhuri.

Naast deze jeugdige bende die om de vijf-zes jaar volledig ververst wordt, is er op MICAS en ESAT ook een ietwat oudere garde die de boel draaiende houdt en van essentieel belang is voor het afleveren van doctoraten: Danielle (de moeder van MICAS), Lut (de moeder van ESAT), Ben (de grote broer van MICAS), Frederik (verlicht despoot van ESAMAT), {Luc, Ludo, Jan en Sven} (The Brotherhood of CDE), {Stef, Piet, Frank, Rik en Mark} (kelderhelden van de systeemgroep), Rudi (voor al mijn productietechnische problemen), Noëlla (professional Bond girl), {Chris en Anne-Marie} (brengers van geld), Thomas, Tony, Michel, Anne, Evelien, Eliane.

Bijna vijf jaar lang heb ik op ESAT in een kelder gezeten. Dat deze kelder slechts half onder de grond zit en er toch wat ramen zijn, verzwijg ik vaak, om de romantiek en het mysterie wat in stand te houden. Desalniettemin blijft het een stinkende kelder (misschien gedeeltelijk door eigen toedoen) en zat ik er soms godganse dagen naar een flatscreen te staren. Dit zou ik nooit hebben volgehouden zonder mijn bureaugenoten Noël en Wouter. Samen met Patrick, maar vaak ook zonder hem, hebben we in ons bureau vele lange technische discussies gehouden die mijn inzichten in de vaak zeer ingewikkelde materie op niet te onderschatten manier hebben aangesterkt. Noël mag gerust een millimeterwavegoeroe genoemd worden en ik heb dan ook vaak beroep kunnen doen op zijn kennis en ervaring om mijn eigen chips te maken. Wouter is technisch natuurlijk ook zeer onderlegd, maar heeft me vaker op niet-technisch vlak nuttige dingen bijgebracht. We hebben niet zelden met een hoop volk rond zijn scherm gestaan om de laatste essentiële YouTube-filmpjes te bekijken of HLN-artikelen te lezen. Tijdens mijn onderzoek heb ik ook veel samengewerkt met mensen uit andere onderzoeksgroepen of bedrijven. Een bedankje dus voor de aangename en vruchtbare samenwerking: Raf, Anton, Cicero en Juan van NXP Eindhoven, Ilja Ocket van Telemic, professor Peter Van Puyvelde van CIT

iv

Natuurlijk is er ook nog leven buiten een doctoraat. Dat leven werd bij mij voor een groot deel opgevuld met muziek. Muziek maken met vrienden is te gek. Het is een eer en een genoegen om dat te mogen doen met Jonas (fig. 1.5 is voor u), Dirk, Jelle, Vincent, Martijn, Jan, Daan, Loes, Silke, Arne, Bert en Gijs bij Carbon, JFJ en Phanques. Mogen luisteren naar andere geniale muziek is ook te gek: Martijn, Ward, Vincent, Jonas, Jan en Jelle van Stad Van Licht, Wartaal en Men of Timber. Naast deze Vrienden van de Muziek zijn er nog andere vrienden, die de rest van mijn tijd opvullen met eten, drinken, films kijken, dansen, samenwonen, sporten en niksen: Michael, Liese, Evy (+ Eden en Aurora), Wim, Sam, Pie, Daan, Gert, Xavier, Annelies, Silvia, Sien, Johan, Koen. Ongeveer precies een jaar geleden kwam ik een meisje tegen dat niet aan mijn dance moves kon weerstaan, maar toch een tijdje het hoofd koel kon houden en pas na lang aandringen mijn lief genoemd mocht worden: Sanne Nijs! Hoewel haar technische bijdragen niet te onderschatten zijn, waren het toch vooral haar softere skills die mij veilig door dit laatste jaar geleid hebben. Dankjewel Sanneke, you’re awesome! En dan is er nog mijn familie, bestaande uit een hoop fantastische mensen met verschillende achternamen: opa en oma Vervaet, opa en oma Tytgat, tantes en nonkels, neefjes en nichtjes. . . . . . en een select groepje met dezelfde achternaam (Tytgat Power!): Annelies, Heleen en Marjon: betere zussen kan een mens zich niet wensen. Papa: voorbeeld en held. En hoewel deze boodschap je nooit zal bereiken, verzend ik ze toch: dankjewel moeke. Maarten Leuven, May 2014

Nederlandse samenvatting

Dankzij de gate-lengteschaling is de analoge snelheid van CMOS-transistoren door de jaren heen significant toegenomen. Dit laat toe om CMOS ook te gebruiken voor millimetergolfsignalen. Deze signalen hebben golflengtes van 1 mm tot 10 mm in vacuüm, wat overeenkomt met frequenties tussen 30 en 300 GHz. De belangrijkste motivatie om op te schuiven naar hogere frequenties is de grote beschikbare bandbreedte. Met deze grote bandbreedte kunnen hoge datasnelheden gehaald worden in telecommunicatieverbindingen, zelfs met eenvoudige modulatieschema’s. Met een bandbreedte van 10 GHz op een draaggolffrequentie van 100 GHz, kunnen bijvoorbeeld signalen met een datasnelheid van respectievelijk 5 Gbit/s en 10 Gbit/s verstuurd worden met BPSK en QPSK. Nog een voordeel van millimetergolven is dat transmissielijnen en zelfs antennes op de chip geïntegreerd kunnen worden dankzij de korte golflengte. Langs de andere kant hebben transistoren een kleinere versterking op hogere frequenties en passieve componenten zoals condensatoren en spoelen vertonen grotere verliezen. Bovendien wordt de free space path loss (FSPL) wel 54 dBm op 100 GHz voor een afstand van slechts 10 cm. Deze zaken wegen zwaar door in het link budget voor telecommunicatietoepassingen. Het uitgezonden vermogen is beperkt, de ontvangen signaal-ruisverhouding (SNR) is laag en het is moeilijk om lage ruisgetallen (NF) te verkrijgen. Gelukkig hoeft de SNR niet erg hoog te zijn om voldoende lage bit error rates (BER) te behalen, dankzij de eenvoudige modulatieschema’s. Dit werk focust zich op het ontwerp van millimetergolfontvangers in CMOStechnologie. Verschillende IC’s werden ontworpen en gemeten met als doel de voorgenoemde problemen aan te pakken. Er wordt in deze thesis eerst een inleiding gegeven over de geschiedenis van (draadloze) telecommunicatie en de evolutie naar millimetergolffrequenties. Daarna worden de bijzonderheden van millimetergolfontwerp in CMOS besproken. De eerste siliciumimplementatie die wordt gepresenteerd, is een 200 GHz downconverter in 90 nm CMOS. Omdat de werkingsfrequentie hoger is dan de fMAX van deze technologie, moet er een passief mixercircuit gebruikt worden. Aangezien er geen low-noise amplifier (LNA) kan gebruikt worden, is de mixer het eerste blok in de ontvanger en wordt het ruisgetal bepaald door het verlies van de mixer. Er wordt een theoretisch kader uitgewerkt om de optimale instellingsspanning en afmetingen van de transistoren in de passieve mixer te bepalen voor minimaal conversieverlies. Zorgvuldige simulaties van zowel de actieve als passieve componenten resulteert in v

vi

een werkende chipimplementatie. Naast de metingen van de ontvangerperformantie op vlak van conversieversterking, lineariteit, ruisgetal en bandbreedte, wordt er ook een testopstelling ontworpen voor het genereren van BPSK- en QPSK-gemoduleerde signalen op 200 GHz. Dit resulteert in oogdiagrammen en constellatieplots van de ontvangen digitale data. Vervolgens wordt het ontwerp van een 120 GHz draadloze verbinding voorgesteld. In dit werk ligt de focus op het ontwerp op systeemniveau, meer in het bijzonder dat van de ontvanger. Een vierfasige Costas-lus (Costas loop) wordt gebruikt om de draaggolf uit het RF-signaal te extraheren en om de twee bits te demoduleren die in de fase zijn gecodeerd. Om voor dezelfde bandbreedte de datasnelheid te vergroten, wordt er een extra bit toegevoegd door het QPSK-signaal ook in amplitude te moduleren. De resulterende constellatie wordt Star-QAM genoemd. De amplitude-informatie kan in parallel met de Costas-lus uit het signaal gehaald worden door middel van een omhullende-detectiecircuit. Er wordt een model gemaakt in MATLAB® om het gedrag op systeemniveau van de volledige link te simuleren en om inzicht te verwerven in de werking van de Costas-lus. Het dynamisch gedrag van de lus en het effect van ruis kunnen zo worden nagegaan. Zo kunnen er ook oogdiagrammen en constellatieplots worden gesimuleerd en kan de verwachte BER berekend worden. In het tweede deel van dit werk worden plastic-golfgeleiderlinks voor millimetergolven onderzocht. Eerst worden de fysische eigenschappen van het kanaal bekeken. Analytische oplossingen voor de electromagnetische golfvoortplanting worden vergeleken met simulaties en metingen. Er wordt gevonden dat voor bepaalde afmetingen van de golfgeleider, de energie meer geconcentreerd is binnen de golfgeleider voor hogere frequenties terwijl de diëlectrische verliezen toenemen. Voor lagere frequenties is de golf minder gebonden aan de golfgeleider en nemen de diëlectrische verliezen af maar gaat er meer vermogen verloren in bochten. Als laatste siliciumimplementatie wordt er een 90 GHz ontvanger in 40 nm CMOS voorgesteld. Het betreft een ASK-ontvanger gebaseerd op injectievergrendeling (injection-locking) met in-chip bonddraadantenna voor een directe koppeling met de plastic golfgeleider. Er wordt een meetopstelling ontworpen om de volledige plastic golfgeleiderlink te testen met een ASK-modulator en de voorgestelde chip. Dankzij het lage verlies in het kanaal, worden signalen ontvangen met een hoge SNR en worden er BER’s bereikt van minder dan 10−12 voor hoge datasnelheden en grote afstanden. De maximale datasnelheid is 9 Gbit/s voor een verbinding van 60 cm en 2.5 Gbit/s voor een verbinding van 9 m. Deze metingen tonen aan dat plastic-golfgeleiderlinks zouden kunnen concurreren met koperverbindingen voor afstanden tot 30 m op vlak van energie-efficiëntie en snelheid.

Abstract

Thanks to the gate length scaling in CMOS, the analog speed of transistors has increased significantly over the years. This has enabled the use of CMOS for millimeter wave signals. These signals have wavelengths between 1 mm and 10 mm in vacuum, which corresponds to frequencies between 30 GHz and 300 GHz. The main motivation for moving to higher frequencies is the large available bandwidth. With this large bandwidth, even with simple modulation schemes, high data rates can be achieved in telecommunication links. For example, with a bandwidth of 10 GHz at a carrier frequency of 100 GHz, signals with a data rate of 5 Gbit/s and 10 Gbit/s can be transmitted using respectively BPSK or QPSK. Another advantage of millimeter wave frequencies is that thanks to the short wavelength, transmission lines and even antennas can be integrated on-chip. On the other hand, at higher frequencies, the transistors have lower gains and passives such as inductors and capacitors exhibit higher losses. Moreover, the free space path loss increases with frequency and becomes 54 dBm at 100 GHz for a distance of only 10 cm. These issues weigh heavily on the link budget for telecommunication applications. The transmitted power is limited, the received SNR is low and low noise figures are hard to achieve. Fortunately, thanks to the rather simple modulation schemes, the SNR does not have to be very high to achieve a low bit error rate. This work focuses on the design of millimeter wave receivers in CMOS technology. Several ICs have been designed and measured in an attempt to tackle the aforementioned problems. In this thesis, first an introduction is given on the history of (wireless) telecommunication and the evolution towards millimeter wave frequencies. Then, the peculiarities of millimeter wave design in CMOS are discussed. The first silicon implementation that is presented, is a 200 GHz downconverter in 90 nm CMOS. Because the operating frequency is higher than the fMAX of the transistors in this technology, a passive mixing circuit must be used. A lot of design effort is put into minimizing the conversion loss of the mixer. Since no LNA can be used, the mixer is the first stage in the receiver and the noise figure is determined by its conversion loss. A theoretical framework is worked out to determine the optimal biasing voltage and sizing of the transistors in the passive mixer for minimum conversion loss. Careful simulation of both the actives and the passives, results in a successful chip implementation. Apart from the measurements of the receiver metrics such as conversion gain, linearity, noise figure and bandwidth, a test setup is also designed to apply BPSK and QPSK modulated signals at 200 GHz to the chip. This results in received eye-diagrams and constellation vii

viii

plots. Next, the design of a 120 GHz wireless link is presented. In this work, the focus is on the system-level design, more in particular that of the receiver. A four-phase Costas loop is used to recover the carrier from the RF-signal and to demodulate the two bits that are encoded in the phase. To increase the data rate for the same bandwidth, an extra bit is added by amplitude-modulating the QPSK signal. The resulting constellation is called Star-QAM. The amplitude information can be retrieved in parallel to the Costas loop in an envelope detection circuit. A MATLAB® model is made to simulate the system-level behaviour of the complete link and to gain insight into the operation of the Costas loop. Loop dynamics, imperfections such as IQ-imbalance and the effects of noise can be evaluated using this model. Eye diagrams and constellation plots can be simulated as well as the expected bit error rate. In a second part of this work, plastic waveguide links for millimeter wave frequencies are investigated. First, the physical properties of the channel are determined. The analytical solutions for the propagation of the EM-field are compared to finite-element simulations and measurements. It is found that for certain waveguide dimensions, the energy is more concentrated inside the waveguide for higher frequencies, while dielectric losses increase. At lower frequencies, the wave is less tightly bound to the waveguide and dielectric losses decrease, but more power is lost in bends. As a last silicon implementation, a 90 GHz receiver in 40 nm CMOS is presented. It is an ASK receiver based on injection-locking, with an on-chip bondwire antenna to interface directly with the plastic waveguide. A measurement setup is designed to demonstrate the plastic waveguide link, containing an ASK modulator and the presented receiver chip. Thanks to the low-loss channel, high SNRs are received and bit error rates lower than 10−12 are reached for high data rates and relatively long lengths. The maximum data rate is 9 Gbit/s for a 60 cm link and 2.5 Gbit/s for a 9 m link. These measurements show that plastic waveguide links might be able to compete with copper links for distances up to 30 m in terms of link energy efficiency and speed.

List of Abbreviations and Symbols

Abbreviations ADC ADSL AGC AM ASK BER BPSK CL CMOS DC DSB-SC DUT EMI EO ESD FIB FDM FM FOM FSK FSPL HDPE IC IIR ILO LAN LFSR

analog-to-digital converter asymmetric digital subscriber line automatic gain control amplitude modulation amplitude shift keying bit error rate binary phase shift keying confidence level complementary metal oxide semiconductor direct current, zero-frequency signal double-sideband suppressed carrier device under test electro-magnetic interference electrical-to-optical electrostatic discharge focused ion beam frequency division multiplexing frequency modulation figure of merit frequency shift keying free space path loss high density polyethylene integrated circuit infinite impulse response injection locked oscillator local area network linear feedback shift register

ix

x

LNA LO MMF MMIC MPSK MOSFET NF NOB NOE OE OOK PA PAM PCB PHM PLL POF POTS PP PRBS PS PSD PSK QAM QPSK QVCO RF RMS Rx SDM SMF SNR SOC SOI SRF TE TEM TDM

low noise amplifier local oscillator monomode fiber monolithic microwave integrated circuit M-ary phase shift keying metal oxide semiconductor field-effect transistor noise figure number of bits number of errors optical-to-electrical on-off keying power amplifier pulse amplitude modulation printed circuit board phase modulation phase locked loop plastic optical fiber plain old telephone system polypropylene pseudo-random binary sequence polystyrene power spectral density phase shift keying quadrature amplitude modulation quadrature phase shift keying quadrature voltage controlled oscillator radio frequency root mean-square receiver space division multiplexing single mode fiber signal-to-noise ratio system on chip silicon on insulator self resonance frequency transverse electric transverse electromagnetic time division multiplexing

xi

TIA TM Tx UTP VCO VGA VNA

transimpedance amplifier transverse magnetic transmitter unshielded twisted pair voltage controlled oscillator variable gain amplifier vector network analyzer

Symbols A B C Cox D E E (ν) Ex y E pq F GA GANT GMAX GP GT H H (ν) Hx IA Ia IIP3 IM3 K KP KVCO L

amplitude, area bandwidth capacitance oxide capacitance symbol or baud rate magnitude of the electric field electric field vector x-component of the electric field in region ν mode in waveguide with main component in y-direction and p extrema in the x-direction and q extrema in the y-direction noise factor (linear), number of fingers in a transistor available power gain antenna gain maximum available gain operating power gain transducer power gain magnitude of the magnetic field magnetic field vector x-component of the magnetic field in region ν DC current AC current in the frequency domain input referred third order intercept point third order intermodulation product stability factor phase detector constant (V/rad) VCO constant (Hz/V) transistor channel length, signal loss, inductance

xii

M N P Pav Pe Q(z) R S Sx y T VA Va Vdd VT W Wf X Z Z∗ Z0 c d e f fMAX fT gds gm h iA ia k l ` r t tox vA

number of symbols in a constellation noise power (signal) power available power probability of error cumulative distribution function of Gaussian distribution resistance, bend radius, bit rate power density per unit area two-port S-parameters (x, y ∈ {1, 2}) temperature, period DC voltage AC voltage in the frequency domain supply voltage MOSFET threshold voltage total transistor width (W = F · W f ) transistor finger width reactance complex impedance complex conjugate of Z characteristic impedance speed of light (3 × 108 m/s), distributed capacitance distance Euler’s number (e ≈ 2.71828) frequency maximum oscillation frequency transition frequency drain to source conductance transconductance Plank’s constant (h ≈ 6.626 × 10−34 Js) total current in time domain (i A = I A + i a ) AC current in the time domain Boltzmann’s constant (1.38 × 10−23 J/K), wavenumber length number of bits per symbol distributed resistance time oxide thickness total voltage in time domain (v A = V A + va )

xiii

va 0 3 α β γ  0 r ζ θ κ λ µ µ0 φ σ ω

AC voltage in the time domain reflection coefficient effect of low-pass filter in time or frequency domain attenuation constant phase constant propagation constant (γ = α + jβ) dielectric constant (permittivity) ( = r 0 ), error signal permittivity in vacuum relative permittivity of a material damping factor phase sensitivity of gds to vG S in a MOSFET in the linear region (κ = ∂gds ∂vG S ) wavelength charge mobility magnetic permeability in vacuum phase error standard deviation pulsation (rad/s) (ω = 2π f )

Table of Contents

Voorwoord

i

Nederlandse samenvatting

v

Abstract

vii

List of Abbreviations and Symbols

ix

Table of Contents

xv

1 Introduction

1

1.1

History of communication . . . . . . . . . . . . . . . . . . . . . . .

1

1.2

Basic principles and trends in electronic communication . . . . . . .

3

1.2.1

Trade-offs in communication systems . . . . . . . . . . . . .

3

1.2.2

Transmitters, channels and receivers . . . . . . . . . . . . . .

4

1.2.3

Advances in microelectronics . . . . . . . . . . . . . . . . .

5

1.2.4

From analog to digital . . . . . . . . . . . . . . . . . . . . .

5

1.2.5

Modulating a carrier . . . . . . . . . . . . . . . . . . . . . .

7

1.2.6

Demodulation . . . . . . . . . . . . . . . . . . . . . . . . . .

9

1.2.7

The sky is not the limit . . . . . . . . . . . . . . . . . . . . .

10

Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

1.3

2 Millimeter wave design in CMOS 2.1

13

Millimeter wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv

13

xvi

TABLE OF CONTENTS

2.1.1

Millimeter wave applications . . . . . . . . . . . . . . . . . .

16

2.2

CMOS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.3

Basic concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

2.3.1

Power and power gain . . . . . . . . . . . . . . . . . . . . .

18

2.3.2

Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

2.3.3

Bit error rate . . . . . . . . . . . . . . . . . . . . . . . . . .

24

2.3.4

Link budget . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

Actives and passives in CMOS . . . . . . . . . . . . . . . . . . . . .

29

2.4.1

Actives . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

29

2.4.2

fT and fMAX . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

2.4.3

Passives . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

32

2.4

3 A 200 GHz down-converter 3.1

3.2

3.3

Circuit architecture and design . . . . . . . . . . . . . . . . . . . . .

38

3.1.1

Design goals . . . . . . . . . . . . . . . . . . . . . . . . . .

39

3.1.2

Noise analysis . . . . . . . . . . . . . . . . . . . . . . . . .

40

3.1.3

Analysis and design of the mixing MOSFETs . . . . . . . . .

41

3.1.4

Transformer design . . . . . . . . . . . . . . . . . . . . . . .

51

3.1.5

Baseband circuits . . . . . . . . . . . . . . . . . . . . . . . .

52

3.1.6

Biasing of mixing transistors and TIA . . . . . . . . . . . . .

53

3.1.7

Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

54

Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

54

3.2.1

Performance metrics . . . . . . . . . . . . . . . . . . . . . .

55

3.2.2

Modulated signals . . . . . . . . . . . . . . . . . . . . . . .

59

Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

4 A 120 GHz star-QAM receiver 4.1

38

System description . . . . . . . . . . . . . . . . . . . . . . . . . . .

66 67

TABLE OF CONTENTS

4.2

4.3

4.4

xvii

4.1.1

The problem of carrier recovery . . . . . . . . . . . . . . . .

67

4.1.2

Star-QAM . . . . . . . . . . . . . . . . . . . . . . . . . . .

69

4.1.3

Link budget . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

4.1.4

System architecture . . . . . . . . . . . . . . . . . . . . . . .

72

4.1.5

The Costas loop for QPSK detection . . . . . . . . . . . . . .

74

4.1.6

Costas loop operation with Star-QAM . . . . . . . . . . . . .

77

4.1.7

Discrete time implementation . . . . . . . . . . . . . . . . .

78

System level simulation . . . . . . . . . . . . . . . . . . . . . . . . .

78

4.2.1

Illustrating the operation with QPSK . . . . . . . . . . . . . .

79

4.2.2

Equivalence to a second order PLL . . . . . . . . . . . . . .

81

4.2.3

Application to the 120 GHz link with Star-QAM . . . . . . .

85

Chip implementation and measurements . . . . . . . . . . . . . . . .

90

4.3.1

The receiver . . . . . . . . . . . . . . . . . . . . . . . . . . .

90

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

90

5 Dielectric waveguides for millimeter waves

92

5.1

Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

5.2

Rectangular plastic waveguides . . . . . . . . . . . . . . . . . . . . .

96

5.2.1

The Marcatili approximation . . . . . . . . . . . . . . . . . .

96

5.2.2

Attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

5.2.3

Dispersion in plastic waveguides . . . . . . . . . . . . . . . . 107

5.2.4

Connector alignment . . . . . . . . . . . . . . . . . . . . . . 109

5.3

Other cross-sections . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

5.4

Dielectric waveguide compared to other channels . . . . . . . . . . . 111

5.5

5.4.1

Comparison between millimeter waves in different channels . 111

5.4.2

Comparison between different high speed wired connections . 113

Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

xviii

TABLE OF CONTENTS

6 An 87 GHz receiver with on-chip bondwire antenna for plastic waveguides

117

6.1

System overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

6.2

Circuit design and layout . . . . . . . . . . . . . . . . . . . . . . . . 119 6.2.1

Low noise amplifier . . . . . . . . . . . . . . . . . . . . . . 119

6.2.2

87 GHz downconversion mixer . . . . . . . . . . . . . . . . . 123

6.2.3

87 GHz injection locked VCO . . . . . . . . . . . . . . . . . 124

6.2.4

Baseband amplifiers . . . . . . . . . . . . . . . . . . . . . . 124

6.3

Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

6.4

Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 6.4.1

Chip characterization . . . . . . . . . . . . . . . . . . . . . . 127

6.4.2

Modulated signals and complete link measurements . . . . . . 130

6.5

Comparison to other work . . . . . . . . . . . . . . . . . . . . . . . 134

6.6

Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

7 Conclusions and future work

137

7.1

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

7.2

Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

List of publications

141

Bibliography

143

Chapter 1 Introduction

1.1

History of communication

If there is one thing in the history of mankind that seems to have improved continuously, it must be the way we communicate. We’ve come from smoke signals and tam-tams via telegraph, telephone and television to online chatting and live video streaming to mobile devices. Figure 1.1 shows the evolution of human communication from the first Homo Sapiens Sapiens around 200 000 BC until today with the most important milestones in history. Our early ancestors developed speech and language around 100 000 BC, which proved to be an enormous evolutionary advantage. Since then, we have found numerous ways to convey information faster, more efficiently, more accurately and over larger distances. After the introduction of speech and language, the first important improvement in communication was the invention of writing around 3000 BC in Mesopotamia. Compared to word of mouth, it improved the reliability of the transmission, decreasing the chance of errors to occur between source and recipient. Note that writing also improves the reliability of data storage, but that is not the topic of this thesis. Pigeon post has been a widely used means of communication from the Roman era until the beginning of the twentieth century, when the telegraph system proved to be faster, cheaper and more reliable. Thanks to the contributions of Alessandro Volta, André-Marie Ampère, Michael Faraday, James Clark Maxwell and many others in the beginning of the nineteenth century, the use of electromagnetism meant a breakthrough in the progress of communication technology. The work that is presented in this thesis is a logical continuation of the developments in electronic communication that started with the experiments of Heinrich Hertz and Guglielmo Marconi with antennas and radio communication. The first transatlantic wireless message was sent and received in 1901. Wireless communication then rapidly evolved to radio and television broadcasting. Wired communication at the same time evolved from telephones over fax to the internet. It is thanks to this evolution that we can talk to relatives on the other side of the globe as if they were in the same room. If you compare the technology of today with the technology of the previous generation (video-chat versus phone booth), you can only try to imagine what will 1

2

1. INTRODUCTION

Figure 1.1: Milestones in communication and electromagnetism. be possible in the next generation. It is probable that the distinction between these different communication networks will fade and that all communication flows through the same network, be it in different physical channels: copper wire, wireless, optical fiber and perhaps one day plastic waveguide. The work presented in this thesis attempts to be a humble contribution to the progress of human communication.

1.2. BASIC PRINCIPLES AND TRENDS IN ELECTRONIC COMMUNICATION

3

Figure 1.2: Trade-offs in communication systems.

1.2

1.2.1

Basic principles and trends in electronic communication Trade-offs in communication systems

Whether a message is carried by acoustic waves, travels hundreds of miles with a horse courier or is being downloaded to our computer, there will always be a tradeoff between speed, reliability and cost. Sometimes, additional conditions have to be taken into account, such as physical size, temperature requirements, security,... This is represented in figure 1.2. When talking about speed, a distinction must be made between the distance traveled per unit of time (propagation speed) and the amount of information transferred per unit of time (data rate). For example, a spoken conversation is limited in data rate but travels at the speed of sound (340 m/s). On the other hand, a horse courier can carry a lot more information at the same time but travels more slowly (about 45 km/s). In a local area network (LAN), data rates of 1 Gbit/s are common. The signals travel at nearly the speed of light (3 × 108 m/s). Time delays are mainly caused by the digital circuitry that processes the signals and are typically below 1 ms. An example of the trade-off between speed and reliability is the mobile network on smartphones. When available, the smartphone will use the 3G or the HSDPA protocol, which allow for high download and upload speeds. But when the network coverage is poor, the phone has to switch to simpler protocols such as EDGE to avoid bit errors, resulting in lower speeds.

4

1. INTRODUCTION

Figure 1.3: A generic communication system consists of a message, a transmitter (Tx), a channel and a receiver (Rx). Evolution in communication means that systems perform better in these three areas: speed, reliability and cost.

1.2.2

Transmitters, channels and receivers

A basic communication link consists of a message, a transmitter, a channel and a receiver. This is shown in figure 1.3. A first restriction made in this thesis is that the message is electronic. It can be an analog waveform or a digital bit sequence, representing real-life data such as audio, video, text or sensor data. Second, the transmitter and the receiver consist of electronic circuits. This excludes optical circuits. Third, the inputs and outputs of the transmitter and receiver are also purely of electrical nature. This excludes optical links (which may possess electronics to process the signals). The output of the transmitter is thus an electrical signal that has to reach the receiver somehow. Now an important distinction has to be made between baseband or passband signalling. The data we want to transfer can be either analog or digital. In both cases, the signals are baseband signals because their spectrum runs from DC to a maximum frequency, called the bandwidth B of the signal. If a signal is not bandlimited by nature, it is usually low-pass filtered. A typical example of an analog data signal is audio. Audible frequencies go from 20 Hz to 20 kHz (but recording goes down to DC). Video signals are also analog. Digital signals can contain any kind of data: text, images, raw sensor data etc. The spectrum of a digital signal depends on the type of baseband filter that is used and is proportional to the data rate of the signal [Cou07, chapter 3]. Baseband signals can be transmitted directly. This is the case in Ethernet LAN or the plain old telephone system (POTS). This is a simple solution but does not allow more than one signal at the same time on the same channel. The receiver would just receive the sum of all the signals on the wire and would not be able to seperate them again. This is one of the reasons to modulate the signal onto a carrier. The resulting signal is then a passband signal. The signal occupies a frequency band around the carrier frequency. The choice between a baseband or a passband system also determines which physical channel can be used. Both can be transmitted through a direct electrical connection. Typically, higher frequencies are attenuated more than lower frequencies, which is why this is generally only used for baseband signals or passband signals with a rather low carrier frequency. For higher carrier frequencies, the attenuation limits the length of

1.2. BASIC PRINCIPLES AND TRENDS IN ELECTRONIC COMMUNICATION

5

the connection. Waveguides that support the transverse electromagnetic (TEM) mode, can be used for both baseband and passband signals because the TEM mode has no lower cut-off frequency. These waveguides must have two or more conductors [Poz05]. This is the case for transmission lines or coaxial cables. Hollow metal waveguides or dielectric waveguides are thus only suited for passband signals. For baseband signals, wireless transmission is restricted to near-field capacitive or inductive coupling, which is also limited in distance. Antennas are band-limited and their size is related to the wavelength of the signal. Therefore, baseband signals can not be radiated by antennas. The circuits in this work are all for modulated, and thus passband, signals. Depending on the application, wireless connections are possible, as well as metal and dielectric waveguides. In all those cases, the electrical signal is converted to an electromagnetic wave which can propagate through the channel.

1.2.3

Advances in microelectronics

Since the invention of the integrated circuit (IC) in 1958, these devices have seen an astonishing evolution. Chips have become smaller, faster and cheaper at the same time. This is all thanks to the improvement of the fabrication process, which enables the production of smaller feature sizes. The effects of scaling were first described by Gordon Moore in his famous paper ”Cramming more components onto integrated circuits“ from 1965 [Moo65]. The smaller gate lengths in the transistors result in lower parasitic capacitance and thus higher speeds. As will be explained in the next chapter, this also means that the highest frequency that can be amplified by the transistors keeps increasing with every new technology node. The most well-known and successful process for integrated circuits is the complementary metal oxide semiconductor (CMOS) process. It is mainly used in digital circuits such as microprocessors and memories but also for image sensors, data converters and highly integrated transceivers. Reduced cost and extensive integration are the main motivators to use CMOS also for high-frequency analog design.

1.2.4

From analog to digital

An electronic signal is called analog when it can vary continuously in time and value. An example is the output voltage of a microphone. As someone speaks into a microphone, the pressure waves that propagate in the air cause a mechanical movement inside the microphone, which in turn causes a change in electrical current or voltage. At each instant, the value of the electrical signal represents the pressure on the microphone. An example of digital signals are the inputs and outputs of a microprocessor. The ones and zeros represent numbers from a finite set. The more bits are used for one number,

6

1. INTRODUCTION

Analog Digital 111

(binary) value

110 101 100 011 010 001 000

time

Figure 1.4: An analog signal and its digitized counterpart the larger this set is and the more detail can be represented. Bits and bytes are read in at discrete time instants, dictated by the digital clock. The higher the clock frequency, the more calculations can be done in a certain time. An analog signal can be digitalized in a so-called analog-to-digital converter (ADC). That means that at predetermined time instants, the value of the analog signal is mapped onto the closest of a limited set of values. The signal is said to be discretized in both time and value. The result is a series of numbers that represent the original signal. Because of the discretisation, this is an approximation of the original signal. The quality depends on the number of possible levels and the speed at which the signal is sampled. Figure 1.4 shows what happens to the analog signal when it is digitized. In this example, three bits are used to represent the eight quantized levels. By choosing the sampling rate and the number of bits high enough, the digital signal can be an arbitrarily accurate version of the analog signal. Of course, a price must be paid in terms of power consumption and cost of the ADC. A limited number of bits results in quantization noise and a too low sampling rate causes aliasing [Cou07]. This analogto-digital conversion happens in every system where real-world analog signals are used to perform digital operations on. In communication systems, there is a trend toward digital systems. Even in the presence of noise, digital signals can be received quasi error-free, because the noise is removed by the digitization if it is small enough. But even digital communication systems need some analog circuits to transmit and receive the signals. These are called the analog front-ends.

1.2. BASIC PRINCIPLES AND TRENDS IN ELECTRONIC COMMUNICATION

1.2.5

7

Modulating a carrier

The carrier is a sine wave with a certain frequency that carries the information by changing one or more of its properties. The unmodulated carrier can be written as: c(t) = A cos (2π f c t + θ ) = A cos (ωc t + θ )

(1.1)

It has an amplitude A, a frequency f c and a phase θ . If a change is induced on these parameters in the transmitter which can be detected in the receiver, information can be conveyed. The modulated signal is called a passband signal, because it occupies a band of 2B around the carrier frequency. Often, this bandwidth is relatively small compared to the carrier frequency, which allows to transmit it using an antenna. By agreeing on using different carrier frequencies for different communication protocols, the full electromagnetic frequency spectrum can be used to transfer data. This is called frequency division multiplexing (FDM). Other forms of multiplexing are time division multiplexing (TDM) and space division multiplexing (SDM). They are often combined to increase the capacitance of a channel. The capacity of the traditional cable for cable TV has been exploited to distribute internet, telephony and radio on the same channel. asymmetric digital subscriber line (ADSL) is a technology that allows to use the POTS for internet with speeds of tens of Mbit/s. In the young history of electronic communications, a clear trend is visible concerning the carrier frequency. Early wireless telegraphs and amplitude modulation (AM) broadcasting worked around 100 kHz to several tens of MHz With the development of frequency modulation (FM) radio, frequencies above 100 MHz. Later, wireless telephony used bands around 0.9 GHz and 1.8 GHz. WiFi works at 2.4 GHz and 5 GHz. The push to higher frequencies comes mainly from the increasing demand for bandwidth. In order to accommodate more and more communication standards and data traffic, the EM-spectrum has to be filled up to ever higher frequencies. The advantage of higher carrier frequencies is the higher absolute bandwidth for the same relative bandwidth. On the other hand, the used electronics limit the maximum frequency that can be processed. Thanks to the improvement in technology, i.e. CMOS scaling, this maximum frequency increases with every new technology node. Analog modulation The analog names for amplitude, phase and frequency modulation are respectively: AM, phase modulation (PM) and FM. Mathematically, PM and FM are equivalent, since changing the phase is the same as changing the instantaneous frequency. Analog modulation is most well-known from analog AM and FM radio broadcast. Analog radio is still very much alive, although AM is already slowly dying, especially

8

1. INTRODUCTION

Figure 1.5: The band Arctic Monkeys named their fifth full-length album from 2013 “AM”. The cover shows an amplitude modulated signal with the letters A and M hidden in the middle. in Europe. Analog TV is not broadcasted anymore on antenna in Belgium since 2011. Analog cable TV is still available but is also expected to be replaced entirely by digital TV in the near future. For more details on analog modulation techniques, the reader is referred to [Cou07, chapter 4]. Digital modulation The digital names for amplitude, phase and frequency modulation are respectively: amplitude shift keying (ASK), phase shift keying (PSK) and frequency shift keying (FSK). ASK and PSK are often combined, which is then called quadrature amplitude modulation (QAM). In this work, only ASK and PSK are used. The general form of an amplitude and phase modulated signal is:
s(t) = A(t) cos 2π f c t + θ (t) (1.2) Where A(t) and θ (t) are derived from the digital baseband signals. The modulated signal s(t) can be represented by a phasor in the complex plane. The real axis is often
called the in-phase component I (t) = A(t) cos θ
(t) and the imaginary axis is called the quadrature component Q(t) = A(t) sin θ (t) . A modulation scheme defines which symbol corresponds to which combination of amplitude and phase, or I and Q. This can be represented in a constellation diagram. The number of constellation points determines the number of bits that are transmitted per symbol. The number of symbols that are transmitted per second is called the symbol

1.2. BASIC PRINCIPLES AND TRENDS IN ELECTRONIC COMMUNICATION

(a) A constellation diagram for QPSK. Each symbol represents two bits.

9

(b) A constellation diagram for 16 QAM. Each symbol represents 4 bits.

Figure 1.6: Examples of constellation diagrams. rate or baud rate, D. It is related to the data rate or bit rate R through the number of bits per symbol ` or the total number of symbols M: R = `D = log2 (M)D

(1.3)

Two examples of constellations are shown in figure 1.6. In quadrature phase shift keying (QPSK), ` = 2 and M = 4, so the bit rate is twice the symbol rate. For 16 QAM, ` = 4 and M = 16, so the bit rate is four times the symbol rate. The receivers in this work are all aimed at digital modulation. They form the analog front-end to a digital communication link.

1.2.6

Demodulation

Demodulation means retrieving the original analog or digital data from the modulated carrier. Two types of demodulators or detectors exist: coherent ones and noncoherent ones. The distinction is made based on the number of inputs. A coherent detector has two inputs: the modulated signal and an unmodulated local oscillator (LO) signal. If the LO frequency has the same frequency as the carrier s(t) and an arbitrary but constant phase θ 0 , multiplication of both signals gives: 
1

 s(t) cos 2π f c t + θ 0 = A(t) cos θ (t) − θ 0 + cos 2π 2 f c t + θ (t) + θ 0 (1.4) 2

10

1. INTRODUCTION

This signal has a baseband component containing the original modulation and a component at twice the carrier frequency which can easily be filtered out by a lowpass filter. If no phase modulation is present, it is clear that the baseband signal is proportional to the original amplitude modulation signal. If no amplitude modulation is present, the baseband signal becomes:
1 1 π
A(t) cos θ (t) − θ 0 = A(t) sin θ (t) − θ 0 + (1.5) 2 2 2 Which acts as a phase detector with a sinusoidal characteristic. A noncoherent detector has only one input: the modulated radio frequency (RF) signal. An example is an envelope detector such as the old and well known crystal radio receiver for demodulation of AM radio. It uses a diode to rectify the signal and thus generates a baseband component. Another option is multiplying the signal with itself in a mixer. At baseband, a signal is generated of which the amplitude is proportional to the RF amplitude squared. Note that phase information can not be retrieved this way.

1.2.7

The sky is not the limit

The Shannon-Hartley theorem states that the maximum data rate or the capacity C of a channel is determined by the received signal-to-noise ratio (SNR) and the bandwidth of the channel: C = B log2 (1 + SNR)

(1.6)

The theorem does not say how this data rate can be reached, it is a theoretical maximum. The concept of “channel” is not restricted to the physical channel in which the signal propagates. For example in the case of wireless transmission, the channel is air. Usually, the bandwidth of air is not the limiting factor in a communication link. Rather, the analog front-end circuits limit the system bandwidth. In some cases, a system specification determines the allowed bandwidth. The SNR of the signal represents the strength of the useful signal power compared to the power of unwanted signals. These unwanted signals can be added to the wanted signal at any stage in the link, even in the electrical circuits themselves, since the components of these circuits generate noise. The bandwidth parameter is the main motivation to use millimeter wave frequencies, as will be explained in the next chapter.

1.3

Outline of the thesis

In this research, several CMOS ICs were designed, fabricated and tested. The main objectives of the research are to demonstrate that. . .

1.3. OUTLINE OF THE THESIS

11

• CMOS is a viable candidate for millimeter wave circuits, despite its weaker performance compared to more specialized RF and microwave technologies, • higher data rates can be achieved by moving to millimeter wave frequencies, • thanks to the reduction in size as a consequence of the higher frequency, transmission lines, transformers and even antennas can be integrated on chip, • cheap plastics can serve as a low-loss channel for millimeter waves, • a link consisting of a CMOS transmitter and receiver and a plastic waveguide can compete with existing wired technologies in terms of data rate and energy consumption. This thesis is structured as follows: Chapter 2 starts with explaining the peculiarities of CMOS and millimeter wave design and how the two interact. It contains some basic concepts that are essential to understanding the rest of the thesis. In chapter 3, the design and measurements of a 200 GHz downconverter are treated. This chip was mainly made to show that CMOS circuits can still work very close to the fMAX of the transistors. As the first chip in this doctorate, it was also an opportunity to learn about the millimeter wave design flow, with the finite-element simulations of passives, parasitic extraction and RF models. We managed to successfully measure conversion gain, bandwidth, linearity and noise figure of the downconverter. At the moment of publication, it was the first CMOS downconverter at such a high frequency with a several GHz bandwidth. Chapter 4 discusses the project that was carried out together with Wouter Volkaerts and Noël Deferm. A 120 GHz wireless connector in 45 nm CMOS with on-chip bondwire antenna was designed and fabricated. The focus of this work is on the system level design of the complete link and in particular the receiver part. The receiver features a four-phase Costas loop for QPSK demodulation and carrier recovery. For this purpose, a computer model was conceived to gain insight in the operation of the Costas loop and to quantify the effect of system parameters on the performance. The second part of this doctorate focuses on the use of CMOS chips in a plastic waveguide link. This is a relatively new area with a lot of unexplored topics. Chapter 5 therefore gives insight in the physical properties of a dielectric waveguide channel. It goes back to the fundamentals of electromagnetism: the Maxwell equations. Theoretical equations for the EM-field are compared with simulations and measurements. The dielectric loss in the waveguide and the bending loss are measured in the lab. Important conclusions can be derived from these measurements concerning the relationship between the dimensions of the waveguide, the frequency and the losses. Then, in chapter 6, a receiver chip is presented which interacts with the plastic waveguide. This injection-locked ASK receiver is measured together with the plastic

12

1. INTRODUCTION

waveguide to reach data rates of several Gbit/s over distances up to 9 m. These measurements also demonstrate the coupling between an on-chip bondwire antenna and a plastic waveguide. Finally, chapter 7 summarizes the achievements in this work and gives some hints towards future work.

Chapter 2 Millimeter wave design in CMOS

This chapter aims to emphasize the particularities of designing millimeter wave circuits in a CMOS technology. Compared to other processes such as SiGe or the III-V compounds GaAs, InP, etc, CMOS is far from the ideal technology for high frequency operation. On the other hand, CMOS allows integration with digital circuits on the same die and is much cheaper for large volumes. The challenge to operate at the limits of the technology’s capabilities and the advantages of CMOS as mentioned before, has encouraged us to design millimeter wave circuits in CMOS. Although the CMOS process has always been optimized for digital circuits, analog designers have always found a way to use those digital transistors to build the analog circuits that make up the front-ends to interact with the real world. About two decades ago, RF-CMOS started booming, allowing the further integration of RF front-ends in CMOS. The carrier frequencies at that time were in the GHz range [Cro95, Kin97, Bor97]. Now, as a positive side-effect of CMOS scaling, the digital transistors are fast enough for analog operation in the millimeter wave frequency range and even the terahertz range. In the first section, the properties of millimeter waves and the opportunities and difficulties are discussed. Some more details of the CMOS process are given. The second section explains some basic concepts that are essential for millimeter wave design and more specific for (wireless) communication links. Power matching, noise, bit error rate (BER) and link budget are discussed in detail. The third section describes the active and passive devices that are used in millimeter wave design in CMOS.

2.1

Millimeter wave

Millimeter waves are electromagnetic waves with wavelengths of 1 to 10 mm in free space. They span a frequency range of 30 to 300 GHz. Figure 2.1 situates the millimeter waves in the electromagnetic spectrum, right in between the infrared and the microwaves. Just above the millimeter wave frequencies are the terahertz frequencies. At this moment, terahertz is the final frontier of electronics. It is the highest frequency range 13

14

2. MILLIMETER WAVE DESIGN IN CMOS

Figure 2.1: The electromagnetic spectrum, with indications of some applications. for which classic electronic circuits are made today. Even higher up the frequency scale, we find the infrared and visible light spectrum. In this range, electronic devices can not provide the gain necessary to build oscillators or amplifiers. Power is generated using lasers, LEDs, or other radiation sources. The main motivation for higher frequencies is the larger absolute bandwidth for the same relative bandwidth. A tuned circuit with a certain quality factor at a low frequency has a smaller absolute bandwidth than a tuned circuit with the same quality factor at a higher frequency. Another feature of millimeter waves is their small wavelength, which is the same order of magnitude as the size of modern integrated circuits. Whereas transmission line effects never occurred on integrated circuits at microwave frequencies, they do

2.1. MILLIMETER WAVE

15

Figure 2.2: Probepads after landing the probes a few times. The pitch is 50 µm. at millimeter wave frequencies. Design techniques for transmission lines on printed circuit board (PCB) can now also be used on chip. Also antennas, inductors and transformers shrink, which allows to integrate them on chip. On the downside, higher frequencies tend to exhibit more loss. The skin effect becomes more pronounced and capacitive coupling to the lossy substrate makes it difficult to make long connections on chip with low loss. Furthermore, measurement setups become more complicated and signal generation is getting more challenging. Above 110 GHz, coax cables are too lossy so rigid metal waveguides must be used instead. Bringing millimeter wave signals on or off the chip is also more difficult. Bondwires typically have an inductance of 1 nH or an impedance of 628  at 100 GHz. The signals are therefore applied or measured using probes. This requires a special probe station and expensive probes that suffer from wear. To demonstrate the destructive effects of probing, figure 2.2 shows the probepads of the 200 GHz downconverter of chapter 3 after several uses. Fortunately in a real application, a chip would not be probed and other solutions exist to transfer millimeter wave signals to and from the IC, such as antenna and flip-chipping. This work presents several chips and their measurement setups. The 200 GHz downconverter was measured on the probe station. The other two chips have bondwire antennas, which solves the problem of probing.

16

2.1.1

2. MILLIMETER WAVE DESIGN IN CMOS

Millimeter wave applications

The applications of millimeter waves can be roughly subdivided into three categories: imaging, radar and communication. Clothing and some organic materials are translucent for some millimeter wave frequencies. This property can be used to determine the maturity of fruit, detect defects in timber or for medical purposes [Miz05, Yuj03]. In airports, security scanners working at 94 GHz can see through clothing to reveal hidden weapons. Radar applications include parking sensors and range detectors for cars [Mit10, Wan08]. The 79 GHz band is allocated for short range car radar systems in Europe. Both the imaging and the radar applications exploit specific properties of the millimeter waves such as transmission, absorption or reflection in certain materials. The information in the received signals is not coming from the transmitted signals (if there are any) but from the objects that are to be detected. This is fundamentally different in communication applications. Here, a transmitter modulates the millimeter wave with digital or analog data and the receiver recovers this data by demodulating the signal. It is this type of application that will be studied in this thesis. Commercial telecommunication products already exist in the millimeter wave range such as WirelessHD™and WiGig™, both operating at 60 GHz and supporting wireless data rates of several Gbit/s. Research is ongoing and a lot of publications emerged in the last decade, with integrated circuits operating at 60 GHz [Afs08, Cho09, LaR09, Tom09, Var07, Wan06, Yao07, Zha13], around 90 GHz [Def11b, Hua11, Las08, Tyt13], 120 GHz [Def11a, Def13, Tak10, Vol11, Vol13], above 200 GHz [Gun08, Kal11, Oje11, Tyt12a] and even entering the terahertz domain [Pfe09, Ste13]. The great benefit of using millimeter wave frequencies for data communication is the large absolute bandwidth that enables high data rates with straightforward modulation schemes. The rest of this chapter discusses the challenges that lie in designing millimeter wave circuits in CMOS, but also the advantages.

2.2

CMOS

CMOS is the preferred technology to produce digital circuits such as processors and RAM chips. The process has also continuously been optimized for these kinds of circuits. The most important aspect that has improved, is the minimum channel length of the transistors. This allows the transistors to switch on and off faster, resulting in faster computing. At the same time, the chip area needed for the same functionality keeps decreasing. This trend is known as CMOS scaling and was first described by Gordon Moore in his famous paper ”Cramming more components onto integrated circuits“ from 1965 [Moo65].

2.3. BASIC CONCEPTS

17

Figure 2.3: A general RF wireless link.

A fortunate side-effect of this scaling is the increase in analog speed, defined by f T and fMAX as described in section 2.4.2. That being said, CMOS is far from the ideal technology for RF and millimeter wave circuits [Doa05]. The III-V semiconductors such as InP and GaAs have a better performance and are historically used to design monolithic microwave integrated circuits (MMIC). These technologies have a higher electron mobility, higher breakdown voltage and high quality-factor passives thanks to their high substrate resistivity. This makes their transistors faster, enables higher output powers en less loss. These technologies however can not offer the same level of integration and low cost as CMOS. In section 2.4.1, the behaviour of metal oxide semiconductor field-effect transistors (MOSFET) at millimeter wave will be described more in depth.

18

2. MILLIMETER WAVE DESIGN IN CMOS

Figure 2.4: A general representation of a cascade of networks, represented by their Thévenin equivalent.

2.3

Basic concepts

A general wireless digital communication system is illustrated in figure 2.3. In the transmitter, the digital baseband signal is converted to the analog domain and upconverted to the RF frequency. The power amplifier (PA) amplifies the signal to drive the antenna. Loss is introduced in the channel and noise is added at the input of the receiver. The low noise amplifier (LNA) amplifies the attenuated signal, the signal is down-converted and translated back to the digital domain. The quality of the received signal determines how many errors are in the demodulated bit stream. This section describes how power gain in the LNA and the PA is defined, how noise degrades the signal and how this limits the maximum transmission distance and data rate.

2.3.1

Power and power gain

At high frequencies, RF power becomes a scarce good and it is easily lost in parasitics. As will be seen, amplifying RF signals becomes more and more difficult as the frequency increases and consumes precious DC power. Therefore, it is important to have a good understanding of the power flows in a millimeter wave circuit. Typically, RF systems can be divided in three parts: the source, the network (which can be a cascade of networks) and the load. In a cascaded network, the input of the next network loads the output of the previous as shown in figure 2.4. Considering linear networks, each can be represented by its Thévenin or Norton equivalent. Note that in general, the input impedance of a network depends on the load impedance and the output impedance depends on the source impedance. The amount of power that flows from the source into the network depends on the matching of the source and the input impedances. If Zin = Z s∗ , there is a conjugate

2.3. BASIC CONCEPTS

19

Figure 2.5: A general representation of a two-port network, with source and load. match, and the power transfer is maximized. In this situation, the current into the input is half the short circuit current Is and the voltage across the input impedance is half the open circuit voltage Vs . The maximum power a source can deliver, is called the available power and depends only on the open circuit voltage (Vs ) or the short circuit current (Is ) and the real part (Rs ) of the source impedance: Pav,s =

I 2 Rs Vs2 = s 8Rs 8

(2.1)

Similarly, the available power at the output of the circuit depends on the open circuit voltage (Vout ) or the short circuit current (Iout ) and the output resistance (Rout ): Pav,out =

V2out I2 Rout = out 8Rout 8

(2.2)

In these equations, Vs , Is , Vout and Iout are AC amplitudes, according to the convention in the list of symbols (p. xi). The actual power delivered to the input or to the load depends on the impedance matching. When there is no conjugate match, reflections occur and only part of the available power is delivered. Consider figure 2.5, where the cascade of figure 2.4 is condensed into one network, represented by its S-parameters. Now three definitions of power gain can be given: transducer power gain G T , available power gain G A and operating power gain G P : GT =

Pl Pav,s

(2.3)

GA =

Pav,out Pav,s

(2.4)

GP =

Pl Pin

(2.5)

20

2. MILLIMETER WAVE DESIGN IN CMOS

In these equations, Pl is the power that is actually delivered to the load and Pin is the power that is actually delivered at the input. From these definitions, it is clear that G A depends on the input match, but not on the output match. That’s because Pav,s is just a property of the source but Pav,out is related to Pin . Similarly, G P depends on the output match, but not on the input match. That’s because Pl is determined by the output match, and Pl and Pin have the same dependence on the input match, so this cancels out in the ratio. The transducer power gain G T depends on both the input match and the output match. These equations can be expressed entirely with the S-parameters of the network and the source and load impedances. This is done by first expressing the impedances in terms of reflection coefficients, with respect to a reference impedance Z 0 , which is usually 50 : 0s =

Zs − Z0 Zs + Z0

(2.6)

0l =

Zl − Z 0 Zl + Z 0

(2.7)

0in = S11 +

S12 S21 0l 1 − S22 0l

(2.8)

0out = S22 +

S12 S21 0s 1 − S11 0s

(2.9)

Which shows the dependence of the input impedance on the load impedance and the dependence of the output impedance on the source impedance. Note that when S12 = 0, this dependence vanishes. The network is then called unilateral. The derivation of the power gains in terms of reflection coefficients and S-parameters can be found in [Gon96]. It results in the following expressions: 2 1 − |0s |2 2 1 − |0l | |S | 21 |1 − S11 0s |2 |1 − 0out 0l |2

(2.10)

2 1 − |0s |2 2 1 − |0l | |S | 21 |1 − 0in 0s |2 |1 − S22 0l |2

(2.11)

GA =

1 − |0s |2 1 |S21 |2 2 |1 − S11 0s | 1 − |0out |2

(2.12)

GP =

2 1 2 1 − |0l | |S | 21 1 − |0in |2 |1 − S22 0l |2

(2.13)

GT =

=

This confirms that indeed, G T depends on both Z s and Z l , G A depends only on Z s and G P only on Z l .

2.3. BASIC CONCEPTS

21

The network in figure 2.5 can be an amplifier, such as a PA or an LNA of which the power gain needs to be known. It can also be a passive network such as a transformer or a transmission line, of which we want to know the loss. Even when there is frequency conversion involved, such as in a mixer, the definitions (2.3-2.5) can be used. For example: the available power gain from RF to IF (conversion gain) when a certain LO power is applied. When the network is a linear two-port, the S-parameters describe the input-output behaviour. A figure that is derived from the S-parameters is the maximum available gain GMAX . It is defined as the power gain that the network would have if both the input and output were a perfect conjugate match. This situation is not always easy to achieve, since the output impedance may depend on the source impedance and the input impedance may depend on the load impedance. But when these conditions are met, there is said to be a simultaneous conjugate match. In this case, GMAX = G A = G P = G T . GMAX can be calculated from the S-parameters: GMAX =

p |S21 | (K − K 2 − 1) |S12 |

(2.14)

with K =

2 | − |S 2 | + |1|2 1 − |S11 22 2|S12 S21 |

(2.15)

and 1 = S11 S22 − S12 S21

(2.16)

Note that GMAX is only defined if K > 1 and |1| < 1, which means that the network is unconditionally stable [Gon96]. In that case, no matter which passive load or source impedance is connected, the network will never oscillate. When these conditions are not met, the device might become unstable for certain load and source impedances and GMAX is not defined. GMAX , also labeled MAG, is thus a property of a two-port network and indicates the maximum gain the network can possibly achieve, and this in the case of a simultaneous conjugate match at input and output. Often, the source and load impedances can not be chosen by the designer. When connecting a device under test (DUT) to measurement equipment, source and load impedances are often 50 . In other cases, the source can be an antenna, or the output of a previous block. Passive matching networks can then be used to transform that impedance to the desired impedance [Lee04, Gon96]. In subsection 2.4.3, some examples of implementations of matching networks in CMOS are given. Of key importance is that the resistive losses in the passive network are not larger than the mismatch loss it has to solve.

vn (V)

22

2. MILLIMETER WAVE DESIGN IN CMOS

3

3

2

2

1

1

0

0

−1

−1

−2

−2

−3 0

0.05

0.1

0.15

0.2 t (s)

0.25

0.3

0.35

0.4 0

0.2

−3 0.4

PDF

Figure 2.6: A low-pass filtered white noise signal in the time domain and its probability density function (PDF). B = 1kHz, σ = Vn,rms = 1V (dashed line).

2.3.2

Noise

Noise can come in different colors and shapes. Its origins can be diverse. Unwanted signals can couple into the system, such as spikes on the mains, 50Hz hum, switching of digital circuits that causes peaks on the supply line etc. But even if all these external sources of noise are eliminated, noise is still present. It is inherent to the movement of electrons in the conductors and the charge carriers (electrons and holes) in semiconductors. Every physical conductor with finite conductivity produces noise. This noise is caused by the random movement of thermally agitated electrons. It can be modeled as a voltage source in series with an ideal noiseless resistor of which the voltage is a random variable with a Gaussian distribution. It has a zero mean value and a non-zero standard deviation σ that depends on the temperature T , the bandwidth B and the resistance R: v u tZ 1 +t2 q u √ u 2 σ = Vn,rms = vn = t lim [vn (t)]2 dt = 4kT R B (2.17) t2 →∞

t1

This is the open-circuit root mean-square (RMS) noise voltage This is called thermal noise or Johnson noise. It is band-limited white noise, which means it has a constant power density for all frequencies within the bandwidth. Note then that the noise power that the resistance produces, depends on the bandwidth in which it is measured. Figure 2.6 shows a low-pass filtered white Gaussian noise signal with an indication of the RMS value. Figure 2.7 shows an equivalent circuit for a noisy resistor with a load

2.3. BASIC CONCEPTS

23

Figure 2.7: Model of a noise resistor, with a load connected. connected to it. From (2.1), the available noise power of a resistor can be derived: Pav,n =

V2n,rms Vn2 = = kT B 8R 4R

(2.18)

The available noise power of a resistor doesn’t depend on the value of the resistance itself. The actual transferred power to the load will of course depend on the matching between R and Zload . This formula for noise power is an approximation of Planck’s law for low frequencies and is only valid for f −1dB can be reached through careful design. At the same time, it is attempted to make sure that the G P = −1dB circle encloses the load impedance and the G A = −1dB circle encloses the source impedance. Increasing the spacing between the two inductors increases the leakage inductance and increasing the trace width increases the capacitance. This way, the G A and G P circles can move around on the Smith chart and can vary in size. It takes some trial and error to end up with an adequate design. Examples of this design procedure can be found in chapter 3 and 6. Transmission lines Different topologies of on-chip transmission lines exist. They can be either single-ended or differential and in the latter case, can be implemented with or without a ground plane. The most important parameters of a transmission line are its characteristic impedance Z 0 , the phase constant β and the attenuation constant α. These are determined by the physical dimensions and the electrical characteristics of the conductors and the dielectrics. Losses are caused by conductor resistance and the conductivity in the oxides and the silicon bulk material. The effect of the physical dimensions on Z 0 and β can be approximated by the lossless case. If the transmission line has a series inductance per unit of length Ls and a parallel capacitance per unit of length of Cp , these parameters become: s Ls (2.40) Z0 = Cp p β = ω Ls Cp

(2.41)

Figure 2.15 shows three common on-chip transmission lines. The capacitance Cp is increased by decreasing the distance between the lines and to the ground plane or by increasing the area of the line(s). The inductance Ls is increased by increasing the distance between the lines and to the ground plane and decreasing the area of the line(s). Of course, when applying these techniques, care must be taken not to increase the series resistance too much. The transmission lines are simulated in a finite-element EM-solver. The resulting twoport parameters can be used to calculate the characteristic impedance and propagation constant. For a lossless transmission line of length l, with a load impedance Z L , the input impedance is: Zin = Z 0

Z L + j Z 0 tan (βl) Z 0 + j Z L tan (βl)

(2.42)

2.4. ACTIVES AND PASSIVES IN CMOS

35

(a)

(b)

(c)

Figure 2.15: Single-ended (a) and differential transmission lines with(b) and without (c) ground plane. When the load impedance is zero, the input impedance is equal to the hybrid parameter H11 and when the load impedance is infinite, the input impedance is equal to the impedance parameter Z 11 . Inserting this in (2.42), we get: Z 11 =

− j Z0 tan (βl)

H11 = j Z 0 tan (βl)

(2.43) (2.44)

So that: Z0 = tan (βl) =

p

Z 11 H11

(2.45)

s

H11 Z 11

(2.46)

These formulas are approximations because in reality, the input impedance will have a real part due to the resistive loss in the transmission line. But for short lengths, the real part is much smaller than the imaginary part. Antennas Antennas are an essential part of any wireless link. They are the interface between the electronic circuit and the electromagnetic wave that propagates through the air. In this work, they are also used to couple electromagnetic energy between a plastic waveguide and a chip (see chapter 6). The antenna converts an electromagnetic wave into a voltage and current or vice versa. This conversion from and to an electromagnetic wave is associated with a certain radiation resistance Rrad . This resistance models the loss of energy that occurs in the conduction electrons when they radiate energy in the transmit antenna. In the receive antenna, this is the internal resistance of the Thévenin equivalent of the ideal antenna.

36

2. MILLIMETER WAVE DESIGN IN CMOS

(a) Transmit and receive antenna.

(b) Circuit equivalent.

Figure 2.16: The circuit equivalent of a transmit and receive antenna. In an ideal antenna, all power is converted between electrical power and radiated power so the efficiency is 100%. In a real antenna however, ohmic losses will be present. They can be modelled as an extra resistance Rloss . Depending on the design of the antenna, the impedance seen in the antenna can have an imaginary part Xant . The total antenna impedance then becomes: Zant = Rrad + Rloss + jXant = Rant + jXant

(2.47)

This is shown in figure 2.16. The open circuit voltage Vant of the receive antenna is determined by the received power. The power delivered to the receiver circuit equals V2

the available power Pav,ant = 8Rant only when the impedances are matched. ant The amount of power that is received depends on the directivity functions of both antennas and the distance between them. Generally, an antenna does not transmit (or receive) power equally in all directions. The directivity function describes the dependency of the transmitted (or received) power in function of the space coordinates. The gain of an antenna is defined as the ratio of power transmitted (or received) in one particular direction to the power transmitted (or received) by a lossless isotropic antenna in that same direction. See also 2.3.4. In certain applications, an omnidirectional antenna is preferred. For example when the location of the transmitter and/or receiver varies. In other cases, when the transmitter and receiver are in fixed locations, directional antennas are preferable, since this results in higher SNR. Directivity is generally achieved by increasing the physical size of

2.4. ACTIVES AND PASSIVES IN CMOS

37

the antenna, using antenna arrays [Joh84]. In a wireless link, even in the absence of interferers, noise will be picked up by the receiver antenna. That’s a consequence of the black-body radiation of objects surrounding the antenna. The power spectral density (PSD) of this noise depends on the temperature of the object. When an antenna is pointed towards space, the perceived temperature is about 4 K. In our applications however, which are mainly indoor, the antenna is surrounded by elements that are assumed to be at room temperature (290 K). In this case, the antenna is indistinguishable from a resistor at temperature T and the available noise is given by Nav,in = kT B

(2.48)

When there are ohmic losses in the antenna, extra noise is added. This can be represented by an extra noise figure of the antenna, which is equal to the power loss. Due to the cost of silicon area, on-chip antennas must be limited in size. The simplest antenna is a half-wave dipole, consisting of two conductors on one line, each of a quarter wavelength long. Thanks to the small wavelength of millimeter waves, these antennas have a size in the same order of magnitude as the chip. The large conductivity of silicon however, results in a small efficiency of dipole antennas in the metal layers of the CMOS technology. The antennas in this work are therefore pseudo-integrated bondwire dipole antennas. In [Def14], the simulated performance of fully integrated and bondwire antennas is compared. It is concluded that bondwire antennas have a better efficiency, higher directivity and larger bandwidth than fully integrated antennas.

Chapter 3 A 200 GHz down-converter

Is it possible to make circuits in CMOS close to or above fMAX ? With this question in mind, the design was started of a down-converter that works at 200 GHz in a 90 nm CMOS technology in which the fMAX of an nMOS transistor is just slightly above 200 GHz. To achieve this, the mixing action in the transistors is based on passive mixing. No gain is required and the transistors don’t have to switch completely on and off at a frequency of 200 GHz. The same principle was already used in 1995, at a frequency of 1.5 GHz [Cro95]. This chapter is structured as follows: Section 3.1 starts with identifying the main design goals and an in-depth noise analysis of the down-converter. Then a thorough analysis of the MOSFET in the linear region and its operation as a linear mixer follows, which allows to choose the design parameters for minimum noise figure. Next, the design of the 200 GHz transformers and of the baseband circuitry are discussed. Measurement results are given in Section 3.2, starting with important performance metrics and proceeding with a demonstration of Gbit/s BPSK and QPSK down-conversion. Section 3.3 concludes with a comparison to publications in the same research area and a summary.

3.1

Circuit architecture and design

The architecture of the complete down-converter chip is shown in figure 3.1 [Tyt11a, Tyt12a]. The core of the circuit is an active MOSFET-C down-converter [Cro95], consisting of four mixing transistors and a transimpedance amplifier (TIA). To achieve passive mixing behavior at 200 GHz, the MOSFETs are used as variable resistors. They are biased in the linear region (VDS = 0 V) and are arranged as a double balanced structure. The LO is applied to the drains and the RF is applied to the gates. This way, a parasitic baseband signal is avoided in the presence of mismatch [Cro95]. The gate voltage varies the MOSFET’s channel resistance, resulting in a current that is proportional to the product of the LO and the RF voltages, as can be seen from the well 38

3.1. CIRCUIT ARCHITECTURE AND DESIGN

Presented work

39

Mixing Transistors

TIA

V-gain

Buffer

Rf

IF+

G VDC S

50Ω

G

Rf

LO 200 GHz

IF-

VG

50Ω

Vdd

200 GHz RF

out-

out+

in+

in-

Vb G

S

G

Figure 3.1: Circuit diagram of the complete 200 GHz down-converter chip. known equation for a MOSFET in the linear region in strong inversion: i DS = µCox

v2 W [(vgs − VT )vds − ds ] L 2

(3.1)

The double balanced structure cancels the unwanted mixing products, while the input capacitance of the TIA (explicitly shown in figure 3.1) acts as an RF short so that only the IF signal remains [Cro95]. To apply the RF and LO signals, probes are used. On-chip transmission lines and transformers match the impedance of the probes to the inputs of the mixer. The gain stages behind the TIA amplify the IF voltage and the buffers are implemented as source followers to drive the external load.

3.1.1

Design goals

The fMAX of the devices in the 90 nm CMOS process is just slightly above 200 GHz. With the unavoidable connection structures attached to the transistors, the fMAX drops below 200 GHz. It is therefore impossible to put an LNA in front of the mixer. In a conventional receiver, the mixer’s noise figure is suppressed by the gain of the LNA so that the receiver noise figure is mainly determined by the LNA performance. In

40

3. A 200 GHZ DOWN-CONVERTER

Figure 3.2: The down-converter as a cascade of different blocks. All indicated powers are available powers. this design however, the system noise figure is determined by the mixer’s noise figure. Hence, the goal is to minimize the mixer’s noise figure and this automatically means minimizing its conversion loss. To boost the down-converted signals for measurement, voltage gain stages are added, with a total gain of 20 dB. These are designed for maximum bandwidth (simulated 8 GHz), in order to demonstrate Gbit/s data communication at 200 GHz. A third important constraint is a low LO power, considering the difficulty to generate high LO amplitudes at these frequencies. Therefore, the LO power is minimized during the design of the mixer.

3.1.2

Noise analysis

Figure 3.2 shows the down-converter as a cascaded system of the mixing transistors, followed by the TIA, the voltage amplifiers and the source followers. The Friis equation [Fri44] can be used to analyze the contribution of the different blocks to the system noise figure. Let Gmix (< 1) and Fmix be the available power conversion gain and noise factor of the mixing transistors. GTIA and FTIA are the available power gain and noise factor of the transimpedance amplifier. G V and FV are the available power gain and noise factor of the voltage gain stage and the source followers combined. Referring to figure 3.2, let PRF be the available signal power presented to the RF input of the down-converter. Pn,RF is the available noise power from the RF source: Pn,RF = kT B. At the output of the mixing quad, PIF is the available signal power at the IF frequency and Pn,IF is the available noise power. Pn,IF is the sum of the attenuated input noise (Pn,RF Gmix ) and the noise generated within the circuit. Since the attenuation of the mixing quad is large (Gmix ' -30 dB), the contribution of Pn,RF can be neglected and only the noise from the mixing transistors plays a role. This is again just thermal noise. There is no 1/f-noise since VDS = 0 so Pn,IF = kT B. For the noise

3.1. CIRCUIT ARCHITECTURE AND DESIGN

41

Table 3.1: Contribution of different blocks to the total NF Block

NF (dB)

F

G A (dB)

Term in (3.3)

Mixing quad

30

1000

-30

1000

TIA

2

1.6

20

600

V-gain + Buffer

6

4

/

30

Total

32.1

1630

/

/

factor of the mixing quad, this results in: Fmix =

SNRRF PRF 1 = = SNRIF PIF G mix

(3.2)

The noise factor of the complete system is then [Fri44]: F=

FTIA − 1 FV − 1 1 + + Gmix Gmix Gmix GTIA

(3.3)

From this equation, the importance of a low 1/Gmix becomes evident. Not only does it directly contribute to the total noise figure, it also amplifies the noise figure of the subsequent stages. As an example, table 3.1 shows the contributions of the different blocks to the total noise figure with realistic values from simulations. The noise figures are determined at baseband (0 - 3 GHz). This example stresses the importance of a low conversion loss in the mixing transistors and a low noise figure of the TIA. Noise simulations are performed on the final design. Figure 3.3 shows the contributions of the different blocks to the input referred noise power spectral density. It’s clearly seen that the 1/f-noise is caused by the TIA and that beyond the 1/f corner (' 10 MHz), both the mixing MOSFETs and the TIA contribute significantly.

3.1.3

Analysis and design of the mixing MOSFETs

The four mixing transistors are designed for maximum conversion gain and minimum required LO power. In this subsection, the influences of the device size and the gate bias voltage are analyzed. After choosing VDS = 0 V and a minimal gate length L = 90 nm, the only design parameters left are the gate width W and the gate bias voltage VGS . The mixing behavior of the MOSFET in the linear region comes from the modulation of the channel conductance by the gate voltage, resulting in a linear multiplication of the small signal drain-source and the gate-source voltages. As indicated in figure 3.4(a), vrf is the differential voltage at the gates and vlo the differential voltage at the drains (VRF and VLO in the frequency domain). The internal

42

3. A 200 GHZ DOWN-CONVERTER

Mixing quad TIA V−gain Buffer Total

Input referred noise PSD (dBm/Hz)

−110 −120 −130 −140 −150 −160 −170 −180 −190 4 10

5

10

6

10

7

10 fIF (Hz)

8

10

9

10

10

10

Figure 3.3: Simulated input referred noise PSD, showing contributions of different blocks. The mixing quad and the TIA are the main contributors to the noise. voltages in the transistors are indicated in figure 3.5 and can be written in the frequency domain as: Vgs = ±

1 1 1 VRF = ± 3VRF 2 (1 + j2π fRF Rg C g ) 2

1 Vds = ± VLO 2

(3.4)

(3.5)

Where Rg is the gate resistance and C g is the gate capacitance of one transistor. The drain and source contact resistances are neglected. The source voltage is assumed to be 0 V in AC. In the time domain, the internal gate-source voltage is denoted by vGS = VGS + vgs , the drain-source voltage is vDS = VDS + vds and iDS = IDS + ids is the total channel current. We will now write the channel current of the MOSFET as a general function of vGS and vDS . By definition: gm ≡

∂iDS ∂vGS

(3.6)

gds ≡

∂iDS ∂vDS

(3.7)

3.1. CIRCUIT ARCHITECTURE AND DESIGN

(a) Double balanced mixing transistors.

(b) AC equivalent of mixing transistors.

Figure 3.4: Double balanced mixing transistors and AC equivalent.

43

44

3. A 200 GHZ DOWN-CONVERTER

Figure 3.5: Simplified equivalent model of the MOSFET in the linear region. Where the channel conductance gds is modulated by vGS . A Taylor expansion on iDS (vGS , vDS ) around the DC point (VGS , VDS ) gives: ∂gds 1 ∂gm ∂gds 1 iDS 'IDS + gm vgs + gds vds + v2gs + vgs vds + v2ds 2 ∂vGS ∂vGS 2 ∂vDS

(3.8)

By arranging the MOSFETs in a double-balanced configuration and summing the currents of the four transistors (i 1 → i 4 ), the unwanted first and second order terms cancel and only the vgs vds term of each transistor remains. Combining (3.8) and (3.4)

3.1. CIRCUIT ARCHITECTURE AND DESIGN

45

and (3.5) gives: isc = i 1 − i 2 + i 3 − i 4 = [IDS + gm

3vrf 3vrf vlo ∂gds 1 3vrf 2 ∂gm 1 vlo ∂gds vlo + gds + ( ) + + ( )2 ] 2 2 2 2 ∂vGS 2 2 ∂vGS 2 2 ∂vDS

− [IDS − gm

3vrf vlo ∂gds 3vrf 1 3vrf 2 ∂gm 1 vlo ∂gds vlo + gds + ( ) − + ( )2 ] 2 2 2 2 ∂vGS 2 2 ∂vGS 2 2 ∂vDS

+ [IDS − gm

3vrf vlo ∂gds 3vrf vlo 1 3vrf 2 ∂gm 1 vlo ∂gds − gds + ( ) + + ( )2 ] 2 2 2 2 ∂vGS 2 2 ∂vGS 2 2 ∂vDS

− [IDS + gm

3vrf 3vrf vlo ∂gds vlo 1 3vrf 2 ∂gm 1 vlo ∂gds − gds + ( ) − + ( )2 ] 2 2 2 2 ∂vGS 2 2 ∂vGS 2 2 ∂vDS

= 3vrf vlo

∂gds ∂vGS

= 3κvrf vlo

(3.9)

With: ∂gds (3.10) ∂vGS κ is the sensitivity of gds with respect to vGS . The short circuit current isc contains a sum frequency and a difference frequency, so let iif be the difference frequency component of isc . As will be seen later on, the low-pass behavior at the gate, represented by 3, has the same cut-off frequency for any value of W , and the influence of VGS on the gate impedance (Rg , C g ) is negligible. That means that the ratio between VRF and Vgs does not vary with W or VGS , so it is convenient to write: κ≡

|IIF | ∝ κ|VRF ||VLO |

(3.11)

Each factor in (3.11) is a function of W and VGS . Since we are interested in the available power conversion gain, we need to know the available output power, which is determined by IIF and the output conductance. Because of the double-balanced configuration, the output conductance of the mixing quad is equal to the small-signal output conductance of one transistor, gds . The capacitive part of the output admittance can be neglected at the IF frequency. While gds is around 1 mS to 100 mS, depending on the size and on VGS , the drain-source capacitance is around 1 fF to 10 fF. This results in a susceptance of 6 µS to 60 µS at 1 GHz. The available IF power from the mixing transistors is: PIF =

|IIF |2 1 ∝ (κ|VRF ||VLO |)2 8gds gds

(3.12)

46

3. A 200 GHZ DOWN-CONVERTER

Figure 3.4(b) shows the AC equivalent circuit for the mixing transistors. VLO and VRF are the small signal drain and gate voltages at fRF = 200 GHz. Z g and Z d are the impedances at the gates and drains. They are equal to the gate and drain impedance of one transistor. The output is specified at the difference frequency fIF . Formulas (3.6) to (3.12) can be made concrete using the transistor equations for the linear region in weak inversion and strong inversion [Tsi99]. In strong inversion: iDS = µCox

v2 W [(vGS − VT )vDS − DS ] L 2

(3.13)

gds ' µCox

W (vGS − VT ) L

(3.14)

W L

(3.15)

κ = µCox In weak inversion: iDS =

W 0 (vGS −VM )/nφt I e (1 − e−vDS /φt ) L M

(3.16)

gds =

W 0 (vGS −VM )/nφt −vDS /φt I e e L M

(3.17)

W 1 0 (vGS −VM )/nφt −vDS /φt I e e L nφt M

(3.18)

κ=

0 , V In these equations, φt = kT /q. The quantities I M M and n are functions of technology related parameters and the constant source-bulk voltage, as defined in [Tsi99] and can be considered as constants in this analysis. Which one of these sets of equations is applicable, depends on VGS . In figure 3.6, iDS , gds and κ of the same transistor are shown versus VGS . The transition between weak and strong can be clearly seen. The effect of mobility degradation at large gate bias is seen in the lower graph. Equation (3.12) expresses PIF in terms of the RF and LO voltages. In order to analyze the conversion loss and noise figure, the analysis must be made in terms of powers. Therefore, PIF is maximized for constant PRF and PLO . First, VRF and VLO are expressed in terms of PRF , Z g and PLO , Z d . For this analysis, it is assumed that the impedances of the RF and LO sources are perfectly matched to the mixer inputs. In reality, the matching will not be perfect and the matching network will introduce extra losses. This effect will be taken into account in 3.1.4. When a power P is dissipated in the real part of an impedance Z = R + j X , the voltage across the impedance is: s 2P.|Z |2 |V | = (3.19) R

3.1. CIRCUIT ARCHITECTURE AND DESIGN

47

Weak inversion / Strong inversion

log(IDS)

−4 −6 −8 −10

gds (mS)

15 10 5 0

κ (mS/V)

30 20 10 0 0

0.2

0.4 0.6 VGS (V)

0.8

1

Figure 3.6: log (IDS ), gds and κ = ∂gds /∂vGS versus VG S for VDS = 0 V. The transition between weak and strong inversion occurs at VG S ' 0.4 V. Mobility degradation is observed for larger VG S . Which is true for both the RF and the LO voltage. This basically means that if the real part of the gate and drain impedance is small compared to the total impedance, for the same LO and RF power, a higher IF power can be achieved according to (3.12). We are now ready to analyze the effect of gate width and VGS on the available conversion gain Gmix , for a fixed LO power. Influence of gate width W From equation (3.11), with (3.15) and (3.18), it is expected that the IF current will benefit from bigger transistors. But for constant power, VRF and VLO decrease with increasing W because Z g and Z d are inversely proportional to W . Simulations with RF models are shown in figure 3.7. Note that both the real parts and the imaginary

48

3. A 200 GHZ DOWN-CONVERTER

parts scale linearly with 1/W , thus justifying the claim that Rg C g in (3.4) does not vary with W . Then from (3.19): r r PRF PLO ; |VLO | ∝ (3.20) |VRF | ∝ W W Inserting this into equation (3.11) and knowing that κ ∝ W , it follows that IIF is independent of W , for fixed PRF and PLO . Equations (3.14) and (3.17) however, show that the output conductance is proportional to the device width. This is also shown in figure 3.7. The conclusion from equation (3.12) is then that the available IF power is inversely proportional to the device width. PIF ∝

1 for constant LO and RF power W

(3.21)

Note that this result does not depend on VGS or whether the transistors are operated in weak or in strong inversion. Simulations were performed in which PRF = -20 dBm and PLO = -20 dBm were held constant, being the actual powers flowing into the gates and drains, respectively. The total gate width was changed. The expected PIF ∝ 1/W behavior is shown in figure 3.8. It should then be concluded that small transistors give more available output power. Very small transistors are not practical however. Figure 3.9 shows the drain and gate impedance on a 50  Smith chart. For very small transistors, in the high impedance region, the points are close to the edge of the Smith chart. These impedances will result in more power loss, as will be explained in section 3.1.4. For this reason, the gate width is chosen to be 6 µm. Influence of VG S When the size of the mixing transistors is chosen, the DC gate voltage must be optimized for maximum available IF power. VGS works in three ways. First, it directly affects gds , as can be seen from (3.14) and (3.17): small VGS gives higher output resistance. Second, it has an influence on the impedances Z g and Z d . The effect on the gate impedance is negligible, but the drain impedance is largely determined by gds thus affecting VLO . Third, κ is a strong function of VGS . The different factors that determine PIF as in (3.12) are drawn in figure 3.10, normalized and on a decibel scale in order to compare their effects. From this graph, it is clear that it’s mainly κ and the output resistance that determine the behavior in function of VGS and that VGS ' 330 mV is the optimum. Note that for this VGS , the transistors are operated in weak inversion. The relative importance of the different contributions in figure 3.10 may depend on W , and so may the optimal value for the gate voltage.

3.1. CIRCUIT ARCHITECTURE AND DESIGN

49

1/Rg

1/10

1/Xg

1/12.5

1/R

d

1/X

1/16.7

g

g

ds

1/Ω

1/25 1/50 0 1/−50 1/−25 2

4

6 8 10 12 14 16 18 20 Total gate width W (µm)

Figure 3.7: Real and imaginary parts of the gate (Z g = Rg + j X g ) and drain (Z d = Rd + j X d ) impedances at 200 GHz and gds versus total gate width of the mixing transistors (simulations on RF models). VGS = 0.4 V.

−30 simulation ~ 1/W

PIF (dBm)

−32 −34 −36 −38 −40 −42 0

2

4

6 8 10 12 14 16 18 20 Total gate width W (µm)

Figure 3.8: Simulated available IF output power of mixing transistors in function of gate width with constant RF and LO power. VGS = 0.4 V.

50

3. A 200 GHZ DOWN-CONVERTER

Zg Zd

W

Figure 3.9: Z g and Z d for W = 2 → 32 µm. VGS = 0.4 V and L = 90 nm. The black encircled impedances represent the design choice, W = 6 µm.

0 −5

normalized, dB

−10 −15 −20 −25

κ 1/gds

−30

VRF

−35

P

VLO IF

−40 0.2

0.4

0.6 VGS (V)

0.8

1

Figure 3.10: Influence of VGS on the different factors in (3.12). PIF reaches a maximum for VGS = 0.33 V. W = 6 µm.

3.1. CIRCUIT ARCHITECTURE AND DESIGN

51

Figure 3.11: Layout of the 200 GHz transformers. Layout of the mixing transistors The transistors are implemented with 6 fingers of 1 µm each, to reduce gate resistance. In drawing the layout for the mixing transistors and the surrounding connections to the transformers, symmetry is very important in order to maintain balanced differential signals. Furthermore, a lot of care is taken to minimize resistive losses, since every dB of loss adds directly to the noise figure. For accurate knowledge of the loss and Z g and Z d , a parasitic extraction is done for the mixing quad and finite element simulations in Momentum are performed for the connection structures around it.

3.1.4

Transformer design

As described in 2.4.3, on-chip transformers can be used as baluns, as a matching networks and to provide a DC bias via the common-mode node. These three functions are applied on the LO and RF inputs of the mixer. Figure 3.11 shows the layout of the transformers. They consist of a single turn in the primary and secondary coil. The traces are made in the top metal, because it has lowest resistivity and least capacitive coupling to the lossy substrate. The design of the transformers follows the procedure of subsection 2.4.3, based on the G A and G P circles. In figure 3.12 the G P and G A circles of -0.75 dB and -1 dB are drawn for the transformer at the gates of the mixing transistors. The gate impedance that is drawn here, is from simulations with parasitic extraction of the transistors and Momentum simulations of the connection structures. The impedances fall just outside the -1 dB circles, resulting in S11 = -7.7 dB with a 50  reference. The ratio of power delivered to the gates to the available RF power (defined as the transducer gain) is simulated to

52

3. A 200 GHZ DOWN-CONVERTER

GPC (−1; −0.75;) GAC (−1; −0.75;) Zgate 50Ω

Figure 3.12: G P and G A circles of the transformer at the gate. The source and load impedances fall just outside the -1 dB circles. be G T = -1.1 dB. Similarly, the transformer at the drain side leads to S11 = -5.1 dB and G T = -2.7 dB. Note that the G P and G A circles lie closer and closer together when approaching the edge of the Smith chart, as can be seen on figure 3.12. Load impedances (Z d or Z g ) in this region of the Smith chart are more likely to fall outside the lower loss circles. In this design, it has not been possible to design transformers that provide a good match between the smallest transistors (W = 2 → 4 µm) and 50 . Choosing W = 6 µm is the best compromise between optimal Gmix and matching. The transformers in the final lay-out have the following sizes: for the transformer at the gate, the diameter of the inner turn is 36 µm and of the outer turn 48 µm. For the transformer at the drain, the diameter of the inner turn is 46 µm and of the outer turn 60 µm.

3.1.5

Baseband circuits

The transimpedance amplifier converts the IF current from the mixing quad into a differential voltage, which is then amplified by additional gain stages. Source followers finally drive the off-chip 50  load. The amplifier of figure 3.13 is used in the TIA and the voltage gain stages. The pMOSTs have their gates connected to ground and serve as load resistors. The value of this resistor determines the gain and the bandwidth of the amplifier. At the same time, it causes a DC voltage drop, so it must also be chosen in function of the DC current. The number of cascaded stages is a trade-off between

3.1. CIRCUIT ARCHITECTURE AND DESIGN

53

Figure 3.13: Schematic of the baseband voltage amplifier. gain and bandwidth. In the final design, a single voltage amplifier has a gain of 10 dB and a bandwidth of 8 GHz. In the TIA, two voltage amplifiers and a feedback resistor of 10 k are used, resulting in a simulated transimpedance of 65.6 dB, a bandwidth of 8.6 GHz and a noise figure of 1.8 dB. A simulated bandwidth of 6.9 GHz is achieved for the complete baseband chain. With a series inductance of 1 nH at the output, representing the bondwires, the bandwidth reduces to 5.7 GHz. The power consumption of the baseband circuits is 63.3 mW and is also the power consumption of the complete chip, since the mixing transistors do not consume DC power.

3.1.6

Biasing of mixing transistors and TIA

Since for the mixing transistors, VDS = 0 V, the gates of the input pair of the TIA can be biased by applying the wanted DC voltage to the drains of the mixing transistors. This voltage VDC is applied to the common mode point on the secondary turn of the transformer. The TIA is designed in such a way that the DC voltage at its output is also VDC and no DC current flows through the feedback resistors. In order to avoid the bulk effect in the mixing quad, these transistors are laid out in a triple well structure, so that their body voltage can be set to VDC independently from the rest of the chip. This technique avoids the use of complicated biasing circuits.

54

3. A 200 GHZ DOWN-CONVERTER

Figure 3.14: Chip micrograph with indication of the active area.

3.1.7

Layout

The down-converter is fabricated in a 90 nm standard CMOS process. The total chip area is 0.375 mm2 , including bond pads, probe pads and decoupling capacitors. The active area as indicated in figure 3.14 is only 0.04 mm2 . The rest of the area is occupied by decoupling capacitors for the DC lines and electrostatic discharge (ESD) protection devices.

3.2

Measurements

For measurements, the chip is glued to an FR-4 PCB. The 200 GHz LO and RF signals are generated by millimeter wave source modules and applied to the chip through G-band waveguide probes. The IF output is brought off-chip through bondwires, connected to a differential transmission line on PCB. The measurement setup on the probe station is shown in figure 3.15. The total power consumption of the downconverter chip is 63.3 mW from a 1.2 V supply.

3.2. MEASUREMENTS

55

Figure 3.15: Photograph of the measurement setup for the IC on an FR-4 PCB on the probe station.

3.2.1

Performance metrics

For an LO with a frequency of 200 GHz and a power of -14.9 dBm and an RF with a frequency of 200.5 GHz and a power of -16.6 dBm, the 500 MHz IF is measured with a power of -10.0 dBm. The conversion gain is thus +6.6 dB. Figure 3.16 shows that the conversion gain varies linearly with the applied LO power, demonstrating the linear multiplication in the MOSFET-C mixer. The 1 dB compression point for the LO power is measured to be -15.6 dBm. Powers higher than -15 dBm could not be generated with the used source modules. Linearity is evaluated by measuring the first harmonic (PIF ) at 500 MHz and the third harmonic (P3 ) at 1.5 GHz at the output, while increasing the RF power. A two-tone test could not be performed because the G-band source modules can only deliver a single tone at the output. However, the third order intermodulation product (IM3 ) can be calculated from P3 through the equation [San06]: IM3 = P3 + 9.54 dB

(3.22)

56

3. A 200 GHZ DOWN-CONVERTER

10 Conversion Gain (dB)

1dB

5

0

−5

−10 −30

−25

−20 PLO (dBm)

−15

Figure 3.16: The measured conversion gain increases linearly with PLO . 10 0

(dBm)

−10 −20 −30 −40 Measured IF Power Simulated IF Power Measured IM3 Power Simulated IM3 Power

−50 −60 −25

−20

−15

−10 PRF (dBm)

−5

0

Figure 3.17: Measurement and simulation of PIF and IM3 versus PRF , with indication of IIP3 = -5.4 dBm. Figure 3.17 shows measured PIF and IM3 . The input third order intermodulation intercept point is IIP3 = -5.4 dBm. Simulations are added for comparison. Note that the linearity is influenced by the LO power, since the IF power is linearly dependent on the LO power and thus the TIA might be driven into saturation. The IF output bandwidth is measured by changing fRF while keeping fLO fixed at 200 GHz. Figure 3.18 shows that the IF bandwidth is 3 GHz. The bandwidth is limited by the baseband circuits, the bondwires and the transmission line on FR-4. The

3.2. MEASUREMENTS

57

−5

PIF (dBm)

−10

3dB

−15

−20 Measured Simulated −25 8 10

9

fIF

10 (Hz)

10

10

Figure 3.18: The measured IF bandwidth is 3 GHz. simulated IF bandwidth of the chip, including bondwires is 5.7 GHz. The LO frequency can be swept along with the RF frequency in the complete G-band. The resulting conversion gain is shown in figure 3.19. Neither the measured nor the simulated conversion gain are at their maximum at 200 GHz. This is due to the imperfect matching of the transformers at 200 GHz. Within a band of 186 GHz to 212 GHz, the conversion gain stays within 4.5 dB to 7.5 dB. This is shown in figure 3.20. Thanks to the wideband 200 GHz transformers, the circuit can operate in an RF range of 26 GHz. The output noise measurement is shown in figure 3.21. The measurements are performed with a spectrum analyzer, using noise markers. This way, the output noise power is measured taking into account the resolution bandwidth and with corrections for the spectrum analyzer’s log amplifier and detector. The ’average’ function is used in order to come to a more accurate result. In the white noise region, the resolution bandwidth can be chosen up to 10 MHz, but in the pink noise region, the resolution bandwidth must be chosen small enough to capture the variation of the noise over frequency. The simulations are also added for comparison. It is observed that the noise analysis of section 2.2 gives a good match to the measurements. The average noise level in a band from 10 kHz to 3 GHz is -137.5 dBm/Hz. The 1/f noise corner frequency is around 10 MHz. Compared to the 3 GHz bandwidth, the 1/f noise can be neglected. From the

58

3. A 200 GHZ DOWN-CONVERTER

12 Measured Simulated

Conversion gain (dB)

10

8

6

4

2

0 140

150

160

170

180 190 fLO (GHz)

200

210

220

Conversion gain (dB)

Figure 3.19: The measured and simulated conversion gain in the G-band.

7.5

4.5

0

186

200 fLO (GHz)

212

Figure 3.20: The measured conversion gain lies between 4.5 dB and 7.5 dB for LO frequencies between 186 and 212 GHz.

3.2. MEASUREMENTS

59

−100 Noise PSD (dBm/Hz)

Simulated Measured −120

−140

−160 4 10

5

10

6

10

7

10 fIF (Hz)

8

10

9

10

10

10

Figure 3.21: The measured and simulated output noise density. The 1/f noise corner is around 10 MHz, the average noise level in a band from 10 kHz to 3 GHz is -137.5 dBm/Hz. integrated noise, the noise figure can be calculated: NF = 10 · log (

No ) = 29.9 dB GkT B

(3.23)

Where No is the integrated output noise power in the band, calculated from the measured data. G is the measured conversion gain and kT B is the available input noise power. This is a single sideband (SSB) noise figure since it assumes a signal on one side of the LO frequency only. This noise figure is comparable to a MOSFET-C down-converter working at 1.5 GHz [Cro95].

3.2.2

Modulated signals

As a demonstration of the IC’s high-speed data communication capability at millimeter wave frequencies, a test setup was conceived to generate 200 GHz BPSK and QPSK modulated signals to apply to the RF input [Tyt11b]. Generation of 200 GHz modulated signals In measurement environments, millimeter wave signals are typically generated using source modules, which extend the frequency range of microwave signal generators to millimeter wave frequencies. Source modules are essentially frequency multipliers which take an RF signal at the input with frequencies of about 8 to 20 GHz and give an output in the wanted frequency band [OML04, Roh10, Far10, Vir04]. Frequencies

60

3. A 200 GHZ DOWN-CONVERTER

up to 1.5 THz are available by choosing the appropriate source module. There are however some limitations in translating modulated signals to higher frequencies using multipliers. Due to the strong nonlinear nature of the source module, any amplitude information in the modulated signal is severely distorted. Phase or frequency modulated signals can be transferred, be it that their modulation depth is multiplied. This multiplication factor for the phase or frequency is however well defined. Frequency multiplication in a source module is typically done in steps of ×2 or ×3 [OML04]. In each step, the signal is fed to a strongly nonlinear device, generating second, third and higher order distortion components. By filtering, only the second or third harmonic remains. By repeating this and amplifying the signal in between, multiplication factors of 4, 6, 8, 12, 18, 24, . . . can be realized. This enables the generation of frequencies up to 1.5 THz from input frequencies between 8 GHz and 20 GHz. Note that when two tones are applied at the input, very strong intermodulation will occur, with large unwanted components in the passband of the source module. It is therefore required that the input be a single carrier signal. The source module needs a certain minimum amplitude at its input and has a certain damage input level. As long as the input amplitude is within this range, the source module will generate a fixed saturated output amplitude which may vary slightly over the frequency band. Because of the frequency multiplication, if a signal A cos(ωt + θ ) is applied at the input, the output will (ideally) be in the form of B cos[M(ωt + θ )] with M being the frequency multiplication factor. There is no linear relationship between A and B, therefore excluding the possibility of conveying amplitude modulation. Any frequency or phase deviation will be multiplied by the known constant M towards the output. This effect has to be taken into account when generating the baseband data signals. The source module requires at its input a single carrier signal u(t) with a frequency f R F /M (typically 8 to 20 GHz) and phase or frequency modulation of which the deviation is divided by M. The required input signal to the source module is generated by a quadrature up-converter architecture, which can generate any amplitude, phase or frequency modulation scheme, based on the I and Q signals x(t) and y(t). These signals are produced by an arbitrary waveform generator (AWG), derived from the following specifications of the wanted signal: modulation scheme, symbol rate, data bit stream, wanted baseband filtering. Care must be taken when specifying the length of the data stream and the sample rate of the signals. Since the stream will be repeated, the number of symbols must be integer and the sample rate must be an integer multiple of the symbol rate, to avoid artifacts. As shown in Fig. 3.22, the baseband signals are mixed with a local oscillator signal cos(ωLO1 t) using two microwave mixers. This gives the signals x(t) cos(ωLO1 t) and y(t) cos(ωLO1 t), which are combined in a 90◦ hybrid coupler and amplified to produce

3.2. MEASUREMENTS

61

Figure 3.22: Setup for the generation of 200 GHz phase or frequency modulated signals. the bandpass signal: u(t) = x(t) cos(ωLO1 t) + y(t) sin(ωLO1 t) = A cos[ωLO1 t + θ1 (t)]

(3.24) (3.25)

After multiplication in the source module, the result is: v(t) = B cos[MωLO1 t + Mθ1 (t)]

(3.26)

= B cos[ωRF t + θRF (t)]

(3.27)

The complete measurement setup for modulated signals is drawn schematically in figure 3.24. In figure 3.23, the constellation plots are shown of the signals that have to be generated at the ωLO1 = ωRF /12 frequency. Note that it is not a good option to let the AWG produce a modulated signal at a certain IF and then up-convert this using one mixer. The maximum sample rate of 12 GS/s per channel limits the bandwidth to 4.8GHz [Tek09]. To accommodate a 4 Gbit/s signal, the IF frequency can be maximum 2.8 GHz. Up-converting this signal with a single mixer would yield a double sideband signal that is not suited for the source module because of the strong intermodulation that occurs. Filtering out one of the sidebands would require a very sharp filter since the bands are spaced by only 5.6 GHz, while the bandwidth is 4 GHz. The quadrature upconversion allows to benefit from the full bandwidth of 4.8 GHz at baseband per channel. That’s why this approach was preferred.

Q

3. A 200 GHZ DOWN-CONVERTER

Q

62

I

I

(a) BPSK/12

(b) QPSK/12

Figure 3.23: Constellation plot of the signal at the input of the frequency multipliers. The phases are divided by M = 12.

Figure 3.24: Schematic representation of the measurement setup for modulated signals.

3.3. CONCLUSION

63

(a)

(b)

Figure 3.25: Measured eye diagrams for BPSK with data rates of 2 Gbit/s (a) and 4 Gbit/s (b), down-converted to baseband. Measured results The output is measured with a Tektronix DPO72004B oscilloscope with a 20 GHz bandwidth. By choosing fRF = fLO , the baseband BPSK signal can be visualized on the oscilloscope as an eye-diagram. This is shown for bit rates of 2 and 4 Gbit/s in figure 3.25. For fRF 6= fLO , the IF signal can be demodulated in the oscilloscope, producing constellation diagrams. BPSK with a bit rate of 2 Gbit/s at an IF of 2.5 GHz and QPSK with a bit rate of 2 Gbit/s at an IF of 2 GHz are shown in figure 3.26.

3.3

Conclusion

In this chapter, a mathematical analysis is made of the 200 GHz mixing MOSFETs in the linear region. A new small signal parameter κ is introduced, which is the sensitivity of the channel conductance with respect to the gate voltage. The influence of the DC gate voltage VGS and the transistor width W on the conversion gain were derived from the current equations of the MOSFET in strong and weak inversion. Simulations confirm the analysis: for best performance, the MOSFETs are operated in weak inversion and small devices provide lower conversion loss and noise figure. The design of a CMOS down-converter was presented, working between 186 and 212 GHz, with a conversion gain of +6.6 dB, a noise figure of 29.9 dB and an IIP3 of -5.4 dBm, for an LO power of -14.9 dBm at 200 GHz. The output bandwidth is 3 GHz,

64

3. A 200 GHZ DOWN-CONVERTER

(a)

(b)

Figure 3.26: Measured down-converted constellation diagrams for BPSK with data rate of 2 Gbit/s (a) and QPSK with data rate of 2 Gbit/s (b), at different IF frequencies. allowing down-conversion of high data rate phase modulated signals as demonstrated up to 4 Gbit/s. In table 3.2, the presented down-converter is compared to other receivers and mixers above 100 GHz, at the time of publication. The table contains circuits in different technologies, of which some have an fMAX larger than the used RF frequency. Therefore, also LNAs and active mixers are included. This work reports the first silicon based receiver able to achieve Gbit/s data rates at 200 GHz or higher. It proves that CMOS, despite its speed limitations, is a viable candidate for integrated circuits at millimeter wave and even terahertz frequencies.

Blocks

LNA + Mixer

Mixer

Mixer

LNA + Mixer

Mixer

Mixer

Mixer

Reference

[Gun08]

[Nic08]

[Pfe09]

[Oje11]

[Ina11]

[Ina11]

This work

130 185 145 161 186 212

45 nm SOI CMOS 90 nm CMOS

650

0.25 µm CMOS

45 nm SOI CMOS

102

65 nm CMOS

202 230

210 225

0.1 µm GaAs mHEMT

0.13 µm SiGe HBT

RF (GHz )

Process

0 - 3000

-

-

-

0 - 1.6

2000

2000

IF (MHz )

+6.6

-4

-12

+18

-

-4

+2

Conv. Gain (dB )

Table 3.2: Performance summary and comparison to other work

Subharmonic Gilbert Double balanced resistive Active single balanced Double balanced MOSFET-C

Square law

Mixer Topology Subharmonic resistive Active double balanced

29.9 DSB

-

-

13 DSB

68 SSB

22

8.4 DSB

NF (dB )

-14.9

2.8

2.8

0 (110 GHz)

-32.5

+1

(109 GHz)

PLO (dBm )

3.3. CONCLUSION 65

Chapter 4 A 120 GHz star-QAM receiver

After the successful design and measurements of the 200 GHz downconverter of the previous chapter, a more complex project was started: a complete millimeter wave wireless link, or “wireless connector” in CMOS. Because signal generation in the existing CMOS processes at that time was not straightforward (if at all possible) at 200 GHz, the frequency was lowered to a more comfortable 120 GHz. Previous research had already demonstrated the possibility of generating sufficient LO power at this frequency [Vol11] and also a transmitter was previously published [Def11a]. The project is ambitious for several reasons. As was already clear from 2.3.4, the SNR at the input of the receiver of a wireless millimeter wave link decreases rapidly as the distance from the transmit antenna increases. The FSPL over a distance of 10 cm at 120 GHz is 54 dB. Due to the limited RF power that can be generated in CMOS at this frequency, this limits the distance over which a signal can be received with acceptable BER. Highly directional antennas could compensate for the FSPL, but these occupy a large area and can thus not be integrated on chip. To process the high data rate, wideband analog and digital baseband circuits are needed. This requirement can be relaxed by using multilevel signaling, i.e. more bits per symbol. This however poses more stringent requirements on the linearity of the transmitter and/or the complexity of the receiver. To avoid the need for probing the RF signals and to demonstrate the feasibility, an on-chip bondwire antenna was designed. This was an extra challenge since there was little or no experience in the research group. Before the actual circuit design, an extensive system level study was carried out. The focus of this chapter will be on system level design and the design of the receiver which contains a four-phase Costas loop for carrier recovery and QPSK demodulation. The design of the quadrature VCO for both transmitter and receiver was carried out by Wouter Volkaerts and is described in [Vol13]. The rest of the circuit was implemented by Noël Deferm. The transmitter chip is presented in [Def13]. Section 4.1 describes the system architecture, with special attention to the receiver. In section 4.2, MATLAB® simulations show the operation of the Costas loop and its

66

4.1. SYSTEM DESCRIPTION

67

equivalence to a phase locked loop (PLL). Section 4.3 briefly discusses the circuit implementation in 45 nm CMOS.

4.1

System description

The goal is to make a transceiver chip in 45 nm CMOS for a wireless link with the following specifications: • Carrier frequency: 120 GHz • Distance: 10 cm • Data rate: 20 Gbit/s • On chip antenna • Power consumption b) but not well bendable around the y-axis

110

5. DIELECTRIC WAVEGUIDES FOR MILLIMETER WAVES

0

loss (dB)

−5

−10

−15

−20 150 GHz 180 GHz 220 GHz −25 0

1

2

3 4 5 6 7 offset in x−direction, ∆ x (mm)

8

9

10

(a)

0 −5 −10

loss (dB)

−15 −20 −25 −30 −35 −40

150 GHz 180 GHz 220 GHz

−45 0

1

2 3 offset in y−direction, ∆ y (mm)

4

5

(b)

Figure 5.17: Measurement of connector misalignment. a = 1.6 mm, b = 0.65 mm.

5.4. DIELECTRIC WAVEGUIDE COMPARED TO OTHER CHANNELS

111

(if a > b), which limits the flexibility. To solve this, square or circular cross-sections could be considered. In [Yeh89], a ribbon waveguide is proposed, which is a rectangular waveguide with a >> b. It is shown that the propagating modes in this configuration show lower loss than in the case where a = 2b. Depending on the permittivity, the factor r Rdiel of equation 5.26 of a ribbon waveguide could be up to 100 times smaller than that of a circular waveguide. Of course, the flexibility around the y-axis is even more limited. Instead of solid waveguides, hollow waveguides can also be used. The advantage is that most of the power is flowing through air, of which the loss tangent is much lower than of plastic. Hollow pipes are however mechanically less robust. In [Kim13], a millimeter wave data link is presented using hollow circular plastic waveguides. Their measured loss of the waveguide is about 1 dB/m at 60 GHz. Compared to 1.5 dB/m at 100 GHz as measured for the 2.2 mm by 0.9 mm rectangular waveguide, this is not a big improvement. More-over, because of the larger crosssectional area (the diameter for the hollow waveguide is 3.2 mm), it is less flexible. In any case, the chip design does not really depend much on the shape of the waveguide. Once the carrier frequency is decided, any type of waveguide can be used, depending on the application. In data centers for example, loss is a more important consideration than mechanical flexibility or size. For consumer applications, flexibility might be more important, but more loss can be tolerated if the distance is limited to a few meters.

5.4

Dielectric waveguide compared to other channels

In the previous section, it was shown how electromagnetic waves travel through dielectric waveguides and what the expected losses are. In this section, a comparison will be made between different data links. This will give insight in which channel is more appropriate for which application.

5.4.1

Comparison between millimeter waves in different channels

Table 5.3 and figure 5.18 compare the loss of four channels at the same frequency of 100 GHz for increasing length. In the case of a wireless link, only the FSPL is taken into account, since atmospheric losses can be neglected as explained in 2.3.4. This loss is quadratic versus distance. The loss in a coax cable, metal waveguide and plastic waveguide are exponential with distance, and have a fixed dB/m value. For the coaxial cable, a loss of 14 dB/m at 100 GHz is found in the datasheet of a manufacturer [Hub11]. The loss in a metal WR-8 waveguide is found to be 5 dB/m [Mila]. In a

112

5. DIELECTRIC WAVEGUIDES FOR MILLIMETER WAVES

Table 5.3: Loss in different channels at 100 GHz Channel

Typical loss

@ 0.1 m

@1m

@ 10 m

Wireless (air)

72 + 20 log d (dB)

52 dB

72 dB

92 dB

Coax cable

14 dB/m

1.4 dB

14 dB

140 dB

Metal WR-8 waveguide

5 dB/m

0.5 dB

5 dB

50 dB

Plastic waveguide

1.5 dB/m

0.15 dB

1.5 dB

15 dB

120 110 loss at 100 GHz in dB

100 90

Wireless (FSPL) Coax (14 dB/m) Metal waveguide (5 dB/m) Plastic waveguide (1.5 dB/m)

80 70 60 50 40 30 20 10 0 −2 10

−1

10

0

10 distance d in meters

1

10

2

10

Figure 5.18: Loss of different channels at 100 GHz. plastic waveguide, a loss of 2 dB/m is assumed, which is conservative compared to the measured loss from figure 5.14(a). Figure 5.18 shows that the loss in air is much higher than in the other channels up to several meters. From a certain distance, the loss in those other channels becomes extremely high, rendering them useless. Only wireless transmission is then feasible at millimeter wave frequencies. Transmission of data at a speed of 40 Gbit/s has been demonstrated over a distance of 1 km at 240 GHz [Fra13]. In this case, high-gain directional antennas (each +15 dB) were used together with dielectric lenses to focus the beam, in combination with a high performance 50 nm mHEMT MMIC technology [Kal11]. This is necessary to generate high RF power in order to overcome the large FSPL. The gain in horn antennas can be 24 dB per antenna [Milb] or more. In that case, the picture of figure 5.18 alters considerably, since the loss curve for a wireless link should be shifted down by 48 dB. This work however focuses on integrated solutions, where the antenna is part of the chip or at least of the package. High gains are only realizable using a large area. This is too expensive in

5.4. DIELECTRIC WAVEGUIDE COMPARED TO OTHER CHANNELS

113

CMOS. Therefore, antennas are assumed to have a rather low gain. Coax cables are used for measurements up to 110 GHz. They become increasingly expensive as the operating frequency rises. This is due to the strict mechanical specifications such as the spacing between the inner conductor and the shield, the use of a low-loss dielectric and the small size. Also the connectors become very critical and expensive. The high loss and the high cost limit their use up to about 1 meter. Metal waveguides are also popular in lab environments. They exhibit lower loss than coax cables, but also become expensive for high frequencies and they are rigid, requiring special parts such as S-bends, 45 degree bends to connect different pieces of equipment. At each interface, extra losses are introduced.

5.4.2

Comparison between different high speed wired connections

The above comparison was done at the same carrier frequency. Only the propagation medium differed. Now let’s see how a millimeter wave dielectric waveguide compares to other wired connections, such as glass and plastic optical fiber and copper wire. The distribution of the internet largely relies on copper wire. The connection to the home is often either via coax cable or via copper twisted pairs. Wired Ethernet networks nowadays usually depend on unshielded twisted pair (UTP) cables. The Cat 5 UTP standard defines a bundle of four twisted copper pairs with a bandwidth of 100 MHz. It supports 100BASE-TX, also known as Fast Ethernet, which is duplex baseband communication with a data rate of maximum 100 Mbit/s over a distance of 100 m using two copper pairs [Wik13]. Using all four pairs and five level pulse amplitude modulation (PAM), also 1000BASE-T or Gigabit Ethernet is supported for speeds up to 1 Gbit/s over a distance of 100 m. With a Cat 6 UTP cable, having a bandwidth of 250 MHz, bit rates up to 10 Gbit/s are supported (10GBASE-T). Coaxial cables are also used for the distribution of internet and TV. The RG-6 coax cable has an attenuation of 0.06 dB/m at 100 MHz [Wik14a]. The cable is used up to 750 MHz. The spectrum is divided in bands to transmit TV, radio and internet. Bitrates go up to 400 Mbit/s for internet, but the same cable can be used for other kinds of digital communication such as digital video (serial digital interface or SDI) with bit rates up to 3 Gbit/s. The data rate is strongly influenced by the quality of the coax cable and the length. Glass fibers can be subdivided in single-mode and multi-mode fibers. Single-mode fibers have the lowest loss (< 1 dB/km) and are used for the transatlantic connection of the world wide web, with repeaters every 300 to 1000 km. Speeds are typically in the range of 40 Gbit/s. Using wave division multiplexing, the capacity of a single fiber can be increased to more than 1 Tbit/s. Multi-mode fibers have a larger diameter. They are cheaper and more suitable for shorter distances, up to 1 km and the supported speeds are lower, up to 10 Gbit/s.

114

5. DIELECTRIC WAVEGUIDES FOR MILLIMETER WAVES

The alignment in connections for glass fibers is very critical because the diameter of the fiber is so small, especially in mono-mode fibers. This, and the high cost of the fibers themselves, makes glass fiber an expensive technology, not suitable for consumers. A cheaper solution for optical links is plastic optical fiber (POF). Plastic optical fibers are always multi-mode because of their large diameter. The losses are in the range of several dB/m. Typical speeds are around 100 Mbit/s for distances up to 100 m although research has shown speeds up to several Gbit/s and distances up to 1 km Figure 5.19 compares copper links to optical fiber links in terms of distance and bit rate. Only single channels are included, i.e. single copper pairs or single optical fibers. Some Ethernet standards are indicated on the graph, of which the details can be easily found online [Wik13]. In [Sin08], 40 Gbit/s link over 24 m coaxial cable is reported. Also the MOST standard is included, which is a POF network for cars. Boundaries indicate the current state of the art performance of copper links, both twisted pair and coax, monomode fiber (MMF) and single mode fiber (SMF). Some plastic waveguide links from literature are added [Fuk11, Kim13, Tyt13]. The goal of this work is ultimately to increase the bit rate × distance product for plastic waveguide links by at least one order of magnitude in order to compete with MMF in terms of bit rate and distance, while still keeping costs low. Copper wires suffer from electro-magnetic interference (EMI) issues and potentially excessive ground-loop currents [Ava12]. Glass or plastic fiber are immune to EMI since typically, sources of EMI are much lower in frequency: power distribution systems, AC/DC power inverters, motor drives and so on. Neither do wireless links such as WiFi or GSM interfere with optical systems. The minimum bend radius for the mentioned links are comparable. In fiber optics and plastic waveguides, bending introduces losses because the EM-wave escapes. For optic fibers, the minimum bend radius is therefore typically 25 mm. In coaxial cables, bending forces the center conductor to one side, causing disturbances in the impedance which results in mismatch and the accompanying loss. Therefore, the minimum bend radius in coax cables is ten to twenty times the outer diameter of the cable. For an RG-6 cable, this comes down to a minimum bend radius of 70 mm. Copper wires can be bent without extra losses. The only limitation lies in the mechanical properties. Four times the outer diameter is enough, which results in a minimum bend radius of about 20 mm for a Cat 5 cable [Wik14b]. Compared to copper, plastic is much lighter. A Cat 5 cable weighs about 30 g/m, an RG-6 coax cable weighs about 45 g/m, while a POF weighs only 6 g/m and a PP waveguide of 2 mm by 1 mm weighs only 2 g/m.

5.5

Conclusion

In this chapter, the properties of rectangular dielectric waveguides were analyzed. The starting point was Marcatili’s set of approximate equations for the EM-field. They are based on the assumption that most of the power is flowing inside the waveguide. This

5.5. CONCLUSION

115

Copper MMF SMF Plastic waveguide Constant bit rate × distance lines 2

10

100GBASE−LR4 [Sin08]

10GBASE−SR

1

10

[Fuk11]

10GBASE−LR

USB 3.0 [Kim13]

10

F

M

M

[Tyt13] 0

1000BASE−SX

ER

PP O

C

bit rate (Gbps)

40GBASE−SR4

100BASE−TX

F

SM

MOST POF

−1

10

10BASE−5

−2

10

−1

10

0

10

1

2

10

10

3

10

4

10

distance (m)

Figure 5.19: Bit rate versus distance for copper and optic fiber links (after [MIT11]). approximation results in a lower cut-off frequency for every mode. The calculated EM-field agrees well with simulations for frequencies above the predicted cut-off frequency. Simulations however show that in fact, there is no cut-off frequency. Most of the power is flowing outside the waveguide at this frequency. For lower frequencies, the wave is actually no longer guided, but propagating in free space. Therefore, the approximation is no longer valid for frequencies close to the cut-off frequency or lower. The fact that the EM-wave is less tightly bound for lower frequencies has two important consequences. First, the dielectric loss approaches 0 dB, since the largest part of the EM-wave is propagating in free space, with a zero loss tangent. This might seem promising, but the second consequence is that the bending losses increase severely. Both effects were confirmed by measurements. These conclusions have a very important impact on the choice of the carrier frequency for the plastic waveguide link. A higher frequency allows more bandwidth and results in a thinner waveguide that is more flexible, with lower bending loss. On the other hand, dielectric loss increases with frequency, power generation becomes more difficult, as

116

5. DIELECTRIC WAVEGUIDES FOR MILLIMETER WAVES

well as the design of the circuits. A compromise was found at 90 GHz, for which a CMOS receiver was designed, as explained in the next chapter. With the properties of plastic waveguides known, they could be compared to other possible solutions. It was shown that plastic waveguides perform better than wireless links, coax cables and metal waveguides when it comes to RF power loss for distances up to several meters. Moreover they are cheaper than both coax and metal waveguides and much more flexible than metal waveguides. In terms of bitrate × distance product, plastic waveguides can easily compete with copper and they have several advantages over copper such as material cost, weight, EMC and galvanic separation. Compared to optical fiber, plastic waveguides do not yet reach the same bitrate distance product, but it is expected that they will achieve the same performance as monomode fiber in the future. The advantages of plastic waveguides over optical fiber are the ease and cost of installation and the cost of the material. These conclusions are an encouragement to build a complete dielectric waveguide link with CMOS transceivers on each end with on-chip antennas. The first step in this link is the development of a CMOS receiver, as discussed in the next chapter.

Chapter 6 An 87 GHz receiver with on-chip bondwire antenna for plastic waveguides

In the previous chapter, plastic waveguides were introduced as a low-loss millimeter wave channel for distances of several meters. In this chapter, the small wavelength at millimeter wave frequencies is exploited to make a direct interface between a chip and a dielectric waveguide with a cross-section in the order of millimeters. This way, a wired data connection can be set up to transmit gigabits of data per second. This chapter presents the design of a 40 nm CMOS receiver chip with bondwire antenna and measurements of a complete plastic waveguide link with a transmitter built from measurement equipment. The first section describes the plastic waveguide link as a system consisting of a transmitter, a receiver chip and a plastic waveguide. Section 6.2 discusses the design of the circuits on the CMOS receiver chip, which is based on the injection locking mechanism. In section 6.3, the design of the onchip bondwire antenna is discussed. Section 6.4 describes the measurements which characterize the chip and the performance of the complete link. Bit error rates measurements up to 9 m distance are presented.

6.1

System overview

A plastic waveguide link consists of a transmitter and a receiver chip on each end of a dielectric waveguide. Antennas couple the millimeter wave signal from the chip to the waveguide and vice versa (figure 6.1). In [Fuk11], these antennas are implemented off-chip on a seperate substrate. In this work, the antennas are implemented using bondwires so they can be integrated with the chip in the same package. A simplified link budget calculation shows the feasibility of the plastic waveguide link in figure 6.1. Table 6.1 shows the link budget for a length of 1 m and 10 m. Assume a transmitted power PTx of 0 dBm, a coupling loss L c of 10 dB at each side and a dielectric loss L  /d of 1.5 dB/m in the waveguide. The bandwidth B of 10 GHz

117

118

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

Figure 6.1: Plastic waveguide link.

Table 6.1: Link budget for different waveguide lengths Distance d (m)

1

10

Transmitted power PTx (dBm)

0

0

1.5

15

L  (dB) 2 × L c (dB) Received power PRx (dBm)

20 -21.5

Modulation bandwidth B (GHz)

10

Received noise power kT B (dB)

-74

Input SNR (dB) Receiver noise figure NF (dB) Output SNR (dB)

52.5

-35

39

10 42.5

29

6.2. CIRCUIT DESIGN AND LAYOUT

119

determines the received noise power. With a receiver noise figure NF of 10 dB, an SNRout of 42.5 dB or 29 dB remains, which is sufficient to receive BPSK, OOK or QPSK with a BER < 10−12 . y In the previous chapter, the E 11 mode was identified as the fundamental mode in a rectangular dielectric waveguide. The transmitter antenna should therefore be designed to couple as much power as possible into this mode. The receiver is designed for a carrier frequency of about 90 GHz. This is a compromise between several factors. A higher frequency makes the design of the circuits more difficult and results in a higher loss in the waveguide. At lower frequencies, thicker waveguides would have to be used to ensure guided propagation, reducing the flexibility. The plastic waveguide has a rectangular cross-section with a width of about 2 mm and a height of about 1 mm. these waveguides interface easily with the WR-10 (2.54 mm × 1.27 mm, 60 GHz → 90 GHz) and the WR-8 (2.03 mm × 1.02 mm, 90 GHz → 140 GHz) metal waveguides used in the measurement equipment.

6.2

Circuit design and layout

The receiver is based on the injection locking principle [Raz04, Adl73]. When the input power to the receiver is high enough, the injection locked oscillator (ILO) is assumed to run at the same frequency as the transmitted signal and with a constant phase difference. To lock the ILO on the RF signal, the signal from the LNA is split as in [Fuk11]. One part is sent to the injection path, which contains an additional buffer stage. The other part goes to the bottom part of a double balanced Gilbert mixer. The ILO removes the modulation from the RF signal. This is only possible when a carrier component is present in the signal. Possible candidates to be demodulated by this injection locked receiver are OOK and (multilevel) ASK. Pure BPSK will not be demodulated by the receiver, since the phase of the ILO will tilt every time the phase of the RF signal tilts. The receiver consists of an LNA, an injection locked VCO, a double balanced mixer and a series of baseband amplifiers. The schematic is shown in figure 6.2.

6.2.1

Low noise amplifier

The LNA is a cascade of three identical transformer-coupled pseudo-differential common source stages. The layout of the complete LNA, together with the input and output networks is shown in figure 6.4. The schematic of the common-source stage is shown in figure 6.3. Neutralization capacitance is added for reverse isolation and differential mode stability. Increasing the reverse isolation means decreasing S12 and thus decreasing the dependence of the input impedance on the load impedance and of

120

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

Figure 6.2: Presented chip.

6.2. CIRCUIT DESIGN AND LAYOUT

121

Figure 6.3: A pseudo-differential common-source pair with capacitive neutralization, used in the LNA and in the injection path to the VCO.

Figure 6.4: Layout of the LNA, including bondpads for antenna (left of reference plane 1) and stub and power splitter (right of reference plane 2). the output impedance on the source impedance. This eases the matching at the input and the output of each stage. Extra resistors are added in the bias lines to improve common mode stability. The design is similar to that of the amplifier in [Def10]. The common-source transistors consist of 16 fingers of 1 µm finger width and the neutralization transistors have 12 fingers of 1 µm. The channel lengths are minimal. These sizes have been chosen with the following considerations: not too large to minimize DC power consumption. Since the RF power levels are rather low, large devices are not needed to ensure linear behaviour. Smaller devices are however more difficult to match, as was explained in section 3.1.4 The bias voltage is chosen for maximum gain and for minimum noise figure. Fortunately, both have the same optimum, for a gate voltage of 700 mV. The input transformer, between the antenna and the first stage of the LNA, is designed to match the impedance of the antenna together with a piece of transmission line, to the input of the transmission line that is connected to the input of the LNA. This

122

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

22 20

S21 (dB)

18 16 14 12 10 8 6 80

85

90

95 f (GHz)

100

105

110

Figure 6.5: Simulated gain (S21 ) of the LNA. The maximum gain is 20.6 dB at 93 GHz and the 3 dB bandwidth is 13.2 GHz, from 86 GHz to 99 GHz. transmission line has a certain length to increase the distance of the LNA to the antenna. This is done in order to avoid coupling from the output of the LNA to the antenna, which might cause instability. The impedance of the antenna, together with the transmission line, is simulated in a full 3D electromagnetic field solver, HFSS™ . The transformers between stages 1 and 2 and between 2 and 3 match the output of the previous stage to the input of the next stage. Behind the last transformer, a parallel stub is added to match the output impedance of the last transformer to the input of the power splitter. The transmission lines, the transformers, the parallel stub and the power splitter are simulated in Momentum, a 2.5D electromagnetic field solver. The design of the transformers follows the procedure of section 2.4.3. The inner and outer diameter of the transformers are respectively from left to right in figure 6.4: 30 µm and 40 µm, 50 µm and 60 µm, 50 µm and 60 µm, 60 µm and 70 µm. The length of the parallel stub is 104 µm. In designing the stub, a trade-off was made between extra losses introduced by the stub and the improved matching. S-parameter simulations with SpectreRF were performed to verify the performance. For the common-source stages, a parasitic extraction was performed on the final layout. This, in combination with the RF-model of the transistors themselves and the Momentum results for the passive structures, gives the transducer gain curve of figure 6.5 and the noise figure of figure 6.6. The gain at 90 GHz is 19.9 dB and the noise figure is 6.7 dB. The simulated input-referred 1-dB compression point is -16.8 dBm, which is sufficiently high to amplify the received signal linearly for the highest power expected (-20 dBm). The simulated power consumption of the LNA is 30.5 mW. The buffer stage is again a common-source stage as in figure 6.3 but smaller: W = 8µm for the common-source transistors and W = 6µm for the neutralization capacitances.

6.2. CIRCUIT DESIGN AND LAYOUT

123

9.5 9

NF (dB)

8.5 8 7.5 7 6.5 80

85

90

95 f (GHz)

100

105

110

Figure 6.6: Simulated noise figure of the LNA. The minimum NF is 6.7 dB at 90 GHz. At the design frequency of 95 GHz, the noise figure is 6.9 dB.

6.2.2

87 GHz downconversion mixer

The downconversion mixer is shown in Fig. 6.7. It’s a classical Gilbert cell with neutralization capacitance in the bottom pair. The pMOS transistors in the linear region act as resistive loads. The neutralization in the bottom pair of the mixer eases the impedance matching between the output of the LNA and the RF input of the mixer. The design of this mixer differs from the 200 GHz downconversion mixer of chapter 3 in several ways. First, it’s an active mixer, which means that the noise figure is not only determined by the conversion gain (or loss), as was the case with the passive 200 GHz mixer. Second, there is an LNA in front of it. That means that the noise figure of the mixer is less determining for the total noise figure of the receiver. Referring to figure 2.8, even if the mixer has a noise figure of 20 dB, the total noise figure increases by only 1 dB because the gain of the LNA is more than 20 dB. The mixer is therefore mainly optimized for conversion gain and low power consumption and not for noise figure. The conversion gain depends on the applied LO power. At 90 GHz, it is quasi impossible to make the top transistors switch completely because the LO power is too low. Simulations predict a mixer noise figure of 22.5 dB and an input-referred 1-dB compression point of -10.7 dBm. This value is too low for the highest expected receiver power of -20 dBm, because the LNA has a gain of 20 dB. This is because at the moment of design, the loss in the plastic waveguides was expected to be higher than what was eventually measured. The simulated power consumption of the mixer is 3 mW.

124

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

Figure 6.7: The 87 GHz downconversion mixer.

6.2.3

87 GHz injection locked VCO

The ILO is shown in Fig. 6.8. It is an LC oscillator, designed to work at 90 GHz, based on [Vol11], and to lock onto the incoming signal to cover dynamic frequency fluctuations in the transmitter VCO. The control voltage can be adjusted to compensate for design errors that result in static frequency deviations. Simulations show an output power of -9.7 dBm when connected to the mixer. The simulated power consumption of the ILO is 12.7 mW.

6.2.4

Baseband amplifiers

The baseband amplifiers are the same as in [Tyt12a]. The last stage is designed to drive the off-chip 50  load.

6.3

Antenna

The antenna is the interface between the plastic waveguide and the receiver circuit. A dipole antenna is made by two bondwires, because this results in lower loss than a dipole fabricated in the top metal [Joh12]. A plane in the top metal acts as a reflector and shield underneath the antenna. The antenna and shield were simulated in HFSS™ ,

6.3. ANTENNA

125

Figure 6.8: The 87 GHz injection locked oscillator. together with the plastic waveguide. The dimensions of the antenna are optimized for maximum power transfer from the waveguide to the chip at 90 GHz. The distance between the outer bondpads of the antenna on chip is 1.4 mm, which is slightly less than λ/2 = 1.7 mm. Of course the effective length of the bondwires is longer than 1.4 mm. The height of the bondwires can be controlled between 0.2 mm and 1 mm but the optimum is found at 0.7 mm. The total chip area occupied by the antenna and the shield is 1.1 mm2 , which is 52% of the total chip area. The area underneath the shield can however be used to place digital or analog baseband circuits. In a lab setup, the plastic waveguide can be brought close to the antenna, suspended in the air. In a real application, a way must be found to fix the chip to the plastic waveguide. Contact between the waveguide and another dielectric must however be avoided, since this will cause signal loss. A possible solution for this is the use of a metal housing that is placed on top of the antenna, in which the waveguide is inserted. Figure 6.10 shows the simulation setup in HFSS™ of the chip with bondwire antenna, plastic waveguide and package. The shape of the housing was optimized to collect as much power as possible in the waveguide. The housing has been prototyped by 3D printing in titanium. The measurements reported in the next section are all performed without the metal housing. Some measurements were performed with the package and no noticeable difference in performance was registered.

126

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

Figure 6.9: Layout of the buffer, VCO and Gilbert mixer together.

6.4

Measurements

The chip is fabricated in a 40 nm standard CMOS process. It is glued and wire bonded to an FR-4 substrate. A chip photograph is shown in figure 6.11. One end of a 2.2 mm by 0.9 mm polypropylene waveguide is inserted into a metal waveguide to apply the RF signal. The other end is brought close to the bondwire antenna on the chip. This is shown in figure 6.12. The RF signal is generated with an Agilent E8257D PSG combined with an OML S12MS source module. For measurements without modulated signals, the plastic waveguide is connected directly to the output of the source module. When modulated signals are needed, the setup in figure 6.16 is used.

6.4. MEASUREMENTS

127

Figure 6.10: Simulation in HFSS™ of the on-chip bondwire antenna with a metal housing and the plastic waveguide.

6.4.1

Chip characterization

The total power consumption is 50 mW from a 0.9 V supply, which includes the buffers to drive the off-chip 50  impedance of the measurement equipment. According to simulations, the LNA consumes 30.5 mW, the VCO 6.9 mW, the downconverter 3 mW and the baseband circuits 14.4 mW. The total area, including decoupling capacitors, antenna, reflector and bondpads, is 2.1 mm2 . The active area is only 0.21 mm2 . The tuning range of the VCO is measured using a spectrum analyzer with an external mixer to pick up the weak signal from the VCO directly with a horn antenna. The tuning range for a control voltage of -1 V to 1 V is 89.5 GHz to 83.1 GHz, as can be seen in figure 6.13. Once the free running frequency of the VCO is set by the control voltage Vctrl , the actual frequency is determined by the locking behaviour. When the input signal to the receiver is large enough, the VCO will lock on the carrier. The more the RF frequency deviates from the free running frequency, the higher this level must be in order for the VCO to lock. The locking behaviour of the ILO can be examined by applying an RF signal and looking at the IF response on a spectrum analyzer. As long as the ILO is locked to the RF input signal, there’s only a DC voltage at the output, but as soon as lock is lost, the difference frequency between the LO and the RF signal is visible on the spectrum analyzer. For each RF frequency, the input power is lowered until lock is lost. This

128

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

Figure 6.11: Micrograph of the chip with the on-chip bondwire antenna at the bottom. power is marked on figure 6.14. On the y-axis, the power coming out of the plastic waveguide, measured with a power meter, is indicated. From the graph, it is seen that lock is most easily achieved at the free running frequency of the VCO. This way, in a complete link, the control voltages of the VCOs in the transmitter and receivers, can be used to bring both frequencies close to each other. This is necessary to compensate for process variations or unexpected parasitics, which influence the VCO frequency. Once in operation, the injection locking will make sure that the receiver VCO tracks the frequency of the transmitter VCO. Although the free running frequency of the ILO is 88.4 GHz for Vctrl = 0 V, the highest IF response is measured for an RF frequency of 87 GHz. There are several possible explanations for this. It’s possible that the operating frequency of the LNA has shifted to 87 GHz. Alternatively, the limited linearity of the mixer results in a better performance at a frequency where the LO power is lower. Therefore, the measurements reported next, are performed at that 87 GHz frequency. The baseband frequency response (figure 6.15) of the chip is measured by applying an 87 GHz RF signal that is amplitude modulated with a sine wave of which the frequency

6.4. MEASUREMENTS

129

Figure 6.12: Chip wire-bonded to PCB with end of plastic waveguide.

VCO frequency (GHz)

90 89 88 87 86 85 84 83

−1

−0.75

−0.5

−0.25 0 0.25 0.5 VCO control voltage (V)

0.75

Figure 6.13: Tuning range of the VCO.

1

130

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

0

PRF (dBm)

−10 −20 −30 −40 −50 85

86

87

88

89

90 91 fRF (GHz)

92

93

94

95

Figure 6.14: Locking range of the VCO, Vctrl = 0V . PRF is the measured power at the output of the waveguide. −10

PIF (dBm)

−15

−20

−25

−30 1

2

3

4

5 6 fIF (GHz)

7

8

9

10

Figure 6.15: Measured baseband frequency response. is increased. The on-chip mixer down-converts this signal to baseband. The resulting sine wave is measured with a spectrum analyzer. The baseband bandwidth is partly limited by the bondwires and the PCB trace.

6.4.2

Modulated signals and complete link measurements

For the generation of Gbps modulated signals at 90 GHz, the setup of figure 6.16 is used. The millimeter wave signal from the source module is amplified to drive the LO port of a 80-90 GHz mixer. The baseband signal comes either from a Tektronix AWG7122B arbitrary waveform generator (AWG) or an Agilent 81250 parallel bit error rate tester (ParBERT). An attenuator is used to set the desired amplitude level. The power that comes out of the metal waveguide is measured with an Erickson PM4

6.4. MEASUREMENTS

131

Figure 6.16: Measurement setup for data rate testing. A flexible plastic waveguide connects the modulated signal generator to the receiver chip. power meter. Unless stated otherwise, this level is set to 0 dBm, a value that is certainly feasible to be generated with modern CMOS processes [Def11a, Fuk11]. To demonstrate multilevel ASK demodulation, the AWG was used to generate a four level ASK signal. The modulated signal is sent through a PS waveguide of 60 cm. An eye diagram of a 4 Gbit/s signal, demodulated by the chip, is shown in figure 6.17. For BPSK signals, the receiver locks on the LO feed-through that occurs in the transmitter of the measurement setup. The ParBERT was used to test bit error rates. For all BER measurements, a PRBS length of 27 − 1 is used. Different lengths of plastic waveguide are tested to determine the influence of the channel, namely power loss and dispersion, on the signal quality. The results can be found in figure 6.18. The carrier frequency is 87 GHz. When a confidence level of 95% is reached for a bit error rate less than BER < 10−12 [Max10], the measurement is stopped in order to avoid excessively long measurement times. The maximum data rate measured for a BER < 10−12 is 9 Gbps, over a distance of 60 cm. The energy efficiency of the receiver is then 5.6 pJ/bit. For a distance of 9 m, the maximum bitrate measured for BER < 10−12 is 2.5 Gbps. The energy efficiency of the receiver is then 20 pJ/bit. At 9 m, the loss in the waveguide is measured to be 23 dB.

132

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

Figure 6.17: Eye diagram of a four level ASK signal of 4 Gbit/s, demodulated by the chip. The transmitted power is 0 dBm at 87 GHz. The channel is a PS waveguide of 60 cm.

1e0

measured BER

1e−3

9m,2Gbps 1m, 5Gbps 0.6m, 7Gbps

1e−6

1e−9

1e−12 −15

−10 −5 0 PRF at input of the plastic waveguide (dBm)

5

Figure 6.18: Measured BER versus input power to the polypropylene waveguide for different data rates and waveguide lengths at 87 GHz carrier frequency.

6.4. MEASUREMENTS

133

a=−1, BER=3.8721e−06

−1

0

1

a=−0.5, BER=1.1045e−05

−1.2649 −0.6325

0

0.6325

1.2649

a=0.5, BER=0.07865

−1.2649 −0.6325

0 0.6325 Amplitude (V)

1.2649

Figure 6.19: When LO feed-through is present but the average signal power is kept constant, the points in the constellation plot get closer together, resulting in higher BER A note on the LO feed-through At the moment of measuring, it was unknown how much LO feed-through was present on the modulated RF signal. Now, for the same average RF power, the power in the modulation is smaller if the power of LO feed-through is larger. This can be seen in figure 6.19. This shows the constellation plots of BPSK signals with increasing LO feed-through but a constant average power. The parameter a denotes the ratio between the two baseband signal levels. It is clear that the BPSK actually becomes a generalized ASK when a 6= −1. The Gauss curves represent the white noise at baseband at the two signal levels. In the figure, an SNR of 10 dB is represented. The larger the overlap between the two Gauss curves, the higher the BER. The actual probability of error can be calculated as [Cou07, page 481]: s  (A − A )2  1 2 Pe = Q (6.1) 2 4σ0 Where Q(z) is the cumulative distribution function for the Gaussian distribution. A1 and A2 are the baseband signal levels of the ASK constellation. σ0 is the standard deviation of the Gaussian distribution, so σ02 is the noise power. Regardless of the LO feed-through, the same noise power is received. It depends only on the system bandwidth.

134

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES

0

10

a=1 0.8 a=

−2

10

75

9m, 2Gbps 1m, 5Gbps 0.6m, 7Gbps

−4

a=0.75

a=0

a=0.5

−6

10

a=−1

BER

10

−8

10

−10

10

−12

10

0

10

20

30

40

50

SNR (dB)

Figure 6.20: BER versus SNR for different values of a, or different levels of LO feedthrough. Also plotted on the same graph: measured BER curves as in figure 6.18. The SNR is derived from the transmitted power. Using this formula, the BER curves in function of SNR can be plotted for different levels of LO feed-through, as shown in figure 6.20. In the same graph, the measured BER curves of figure 6.18 are repeated. The SNR is computed from the transmitted power, an estimation of the channel loss in the plastic and the NF of the receiver. The NF of the receiver is derived from simulations of the circuits and also includes the loss of the receive antenna. As can be seen, the measured curves correspond to an a factor of about 0.8 to 0.9, which indicates a large LO feed-through.

6.5

Comparison to other work

Table 6.2 summarizes the performance and compares this work to the two receivers in [Fuk11] and the receiver in [Kim13]. In [Fuk11], the authors propose a full-duplex interconnect using two links at different frequencies in the same plastic waveguide. Their maximum link distance is only 1 m but they achieve a higher bit rate of 15 Gbit/s. The antenna they use is rather large with an area on PCB of about 4 mm2 . Their plastic waveguide accommodates two links at the same time and as a consequence has a cross-section of 0.8 by 8 mm, approaching the ribbon-like waveguide as proposed in [Yeh89]. It is therefore only bendable in one direction. Their power consumption is

6.6. CONCLUSION

135

comparable to this work but because of their limited link distance, the figure of merit (FOM) is worse (4.2 pJ/bit/m versus 2.2 pJ/bit/m in this work). The link in [Kim13] achieves a slightly higher bit rate - distance product and a lower power consumption, resulting in a better FOM of 0.64 pJ/bit/m. Their hollow circular waveguide with a diameter of 3.2 mm has a lower loss of 1 dB/m at 60 GHz compared to 1.5 dB/m at 90 GHz used in this work. To date, this work reports the highest carrier frequency and the largest distance in plastic waveguide links.

6.6

Conclusion

In this chapter, the design of an injection-locked ASK receiver in 40-nm CMOS with on-chip bondwire antenna is presented. It was used to demonstrate a plastic waveguide link. A maximum data rate of 9 Gbit/s was achieved for a link distance of 60 cm. For a length of 9 m, the maximum achieved data rate was 2.5 Gbit/s. The energy per bit per meter for plastic waveguide links is in the order of 1 pJ/bit/m [Fuk11, Kim13, Tyt13]. This is one order of magnitude lower than wired electrical links and about one order of magnitude higher than optical links [Kri11]. Figure 5.19 already showed that the distance - bit rate product competes with copper links today at 1000 Gbit/s · m and is expected to evolve to MMF-like performance at 10000 Gbit/s · m. Research on plastic waveguide links using IC transceivers is still very young. A lot of improvement is to be expected in the near future. By optimizing the circuits, power consumption can be lowered further and the bandwidth can be increased. More complex modulation schemes are an opportunity to increase the data rate. This in contrast with optical links, where only ASK is possible, since photodetectors can only measure intensity and not phase. Moreover, no opto-electrical conversion is needed. This reduces the complexity and the cost of the system. Optimization of the plastic waveguide is expected to reduce the dielectric loss. This way, longer distances can be achieved.

6. AN 87 GHZ RECEIVER WITH ON-CHIP BONDWIRE ANTENNA FOR PLASTIC WAVEGUIDES 136

Carrier Frequency Tx power Highest bitrate at distance Bitrate × distance Longest distance Maximum bitrate Bitrate × distance DC power* Best FOM* Supply Modulation Active area* Architecture Antenna Technology Plastic Waveguide

* Receiver only

[Fuk11] 80 GHz Rx 80 GHz 0 dBm 12.5 Gbps 0.12 m 1.5 Gbps.m

45 mW 30 pJ/bit/m 1.1 V ASK 0.14 mm2 Injection locking Quasi-Yagi on PCB 40 nm CMOS Rectangular 8 × 0.8 mm2

60 GHz 0 dBm 6 Gbps 2m 12 Gbps.m 7.6 m 3.3 Gbps 25.1 Gbps.m 16 mW 0.64 pJ/bit/m 1V ASK 0.08 mm2 Self mixing Dipole off-chip bondwire 65 nm CMOS Hollow circular 3.2 mm

[Kim13]

87 GHz 0 dBm 9 Gbps 0.6 m 5.4 Gbps.m 9m 2.5 Gbps 22.5 Gbps.m 50 mW 2.2 pJ/bit/m 0.9 V (multilevel) ASK 0.21 mm2 Injection locking Dipole on-chip bondwire 40 nm CMOS Rectangular 2.2 × 0.9 mm2

This work

Table 6.2: Performance summary and comparison to other work [Fuk11] 57 GHz Rx 57 GHz 0 dBm 15 Gbps 0.1 m 1.5 Gbps.m 1m 10 Gbps 10 Gbps.m 42 mW 4.2 pJ/bit/m 1.1 V ASK 0.14 mm2 Injection locking Quasi-Yagi on PCB 40 nm CMOS Rectangular 8 × 0.8 mm2

Chapter 7 Conclusions and future work

7.1

Conclusions

Throughout history, communication technology has made progress in three important domains: speed, accuracy and cost. Cost can be generalized to production cost, power consumption, and even size. Speed is interpreted here as the amount of data that can be transmitted per unit of time. It is assumed that the travelling time is not an issue since it concerns electromagnetic signals, travelling at nearly light speed over relatively short distances. The capacity, describing the speed or maximum data rate achievable through a channel is determined by its bandwidth B and the received signal-to-noise ratio. In the case of wireless communication, the channel bandwidth is virtually infinite, but the system bandwidth is limited by the analog front-end circuits. In a way, the analog front-end can be seen as part of the channel. The SNR at the point of decision in the receiver, determines the bit error rate and thus the accuracy of digital communication. The link budget identifies the factors that determine this SNR. First of all, the noise figure of the receiver should be as low as possible. Second, the total signal attenuation between the transmitter and the receiver should be minimized, since the maximum output power of the transmitter is limited. This work investigates millimeter waves as a solution to the bandwidth problem. Tuned circuits typically have a relative bandwidth of about 10%. So by increasing the carrier frequency, the absolute bandwidth is also increased. This is the main motivation to go to higher frequencies. The CMOS process is used in all the presented receiver chips. This is motivated by the cost of the system. CMOS allows extensive integration, which makes it possible to include high-performance digital circuits on the same die as the analog RF front-end. This reduces the amount of seperate chips in a system, making devices smaller and more portable. Moreover, production costs decrease, especially for large production volumes. Of course, there must be a catch. Because of the high frequency, losses on the chip increase, which makes it difficult to realize high gains for low DC power. This is reflected in the energy efficiency at both the transmitter and the receiver. On top of 137

138

7. CONCLUSIONS AND FUTURE WORK

that, the free space path loss becomes very large. To ensure a sufficient SNR at the receiver without excessive transmitted power, a solution must be found to overcome this loss. This can be done in two ways: limiting the distance, as is done in the 120 GHz wireless link of chapter 4 or introducing directivity. This directivity can be achieved by using high-gain antennas. These can however not be implemented in CMOS, because this requires a very large area and would be very costly. Off-chip solutions exist, as demonstrated in [Fra13], but these systems are very bulky and expensive. Directivity can also be implemented by using a guided wave. A very elegant solution is provided by dielectric waveguides. As described in chapter 5, polymer waveguides are presented as a low-cost, low-loss and flexible channel to reach distances up to 10 m. In this thesis, the implications of using millimeter waves in combination with the CMOS process are elaborated in chapter 2. First, some basic concepts are explained such as the different definitions of power gain, the relation between noise and BER, link budget and FSPL. The performance of CMOS at millimeter wave frequencies is discussed next, both the transistors and the passive devices such as transformers, transmission lines and antennas. In the 200 GHz downconverter of chapter 3, the frequency boundary of CMOS circuits for data communication is pushed. At the time of publication, it was the only silicon based circuit above 200 GHz able to handle Gbit/s speeds. The theoretical influence of the DC gate voltage and the transistor size on the conversion gain is confirmed by simulations. 200 GHz BPSK and QPSK signals are down-converted with the chip and good eye diagrams and constellation plots are shown for data rates up to 4 Gbit/s. Despite the extremely high frequency and the transistor models being validated only up to 30 GHz, a good match is obtained between the simulations and the measurements of the conversion gain and the noise figure. A more complex system is described in chapter 4: a wireless link at 120 GHz. To cope with the problem of carrier recovery, a four-phase Costas loop is used in the receiver. A computer model for the complete wireless link is made to determine the influence of design parameters on the link performance and to design the system. This model shows the equivalence of the Costas loop to a second-order PLL. It can be used to show the effect of noise and imperfections such as IQ-imbalance. The model is also able to produce eye diagrams and constellation plots and compute BER. In a second part of this doctorate, dielectric waveguide links are investigated. Chapter 5 presents plastic waveguides as a low-cost and low-loss channel for millimeter waves. First, Marcatili’s analytical electromagnetic field equations are compared to finite-element simulations. A good match is found for sufficiently high frequencies, above the cut-off frequency predicted by Marcatili. Simulations show that for lower frequencies, the wave decouples from the waveguide and the Marcatili approximation no longer holds. Bending losses increase but dielectric loss decreases because the wave resembles more and more a plane wave in vacuum. At higher frequencies, the wave is more concentrated in the waveguide, resulting in lower bending loss but higher dielectric loss. These conclusions are also confirmed by measurements on plastic

7.2. FUTURE WORK

139

waveguides. Simulations of the loss also match very well with the measurements. At 100 GHz, a loss of 1.5 dB/m is found in a polypropylene waveguide of 2.2 mm by 0.9 mm. This results in a loss of only 15 dB over 10 m, which is very low compared to the FSPL of 92 dB for the same distance. Finally, in chapter 6, a 90 GHz receiver is presented with an on-chip bondwire antenna as an interface with the plastic waveguide. The receiver is based on the injection-locking mechanism and works for any type of ASK modulation. It is tested with modulated signals and with a BER < 10−12 , a maximum data rate of 9 Gbit/s is achieved over 60 cm and 2.5 Gbit/s is achieved over 9 m. This results in an energy efficiency per unit of distance of 2.2 pJ/bit/m, which is an order of magnitude better than typical copper links.

7.2

Future work

The chips in this thesis have been designed for high-speed data communication. All three designs serve the purpose of receiving millimeter wave signals with a large bandwidth. The specifications depend largely on the link budget which is in turn determined by the transmission channel and the distance. In this work, two different channels were targeted: wireless through air and wired through a plastic waveguide. In both configurations there is still room for improvement on the system level, on the circuit level and on the physical side (antenna, waveguide, . . . ), in order to achieve 20 to 30 Gbit/s over larger distances. In general, the performance of CMOS millimeter wave circuits can be improved by using more accurate transistor models. The models used in this work are not verified by measurements up to the frequencies at which they are used, but typically only up to 30 GHz. Therefore, the impedance matching may not be perfect and power gains are lower than they could be. Moreover, using the RF-models provided by the foundry, limits the changes that can be made in the layout. The layout of the transistors can therefore not be optimized for millimeter wave performance. In this work, only ASK and PSK and a combination thereof have been used. It can be useful to look into FSK as well and investigate whether it could be more power-efficient or result in simpler architectures. In both wireless and plastic waveguide links, improvements can be expected in the design of the antenna. The antennas used in this work are not completely integrated, because the performance of a bondwire antenna is better than that of a fully-integrated dipole. However, alternative packaging solutions such as flip-chip bonding could allow the use of more integrated antenna configurations with similar performance. In some applications, a highly-directive antenna may be required. This can be realized as a phase-array, but occupies a large area. It may therefore not be realistic to integrate it on-chip but look for alternative solutions such as horn antennas or planar antennas on microwave substrates.

140

7. CONCLUSIONS AND FUTURE WORK

In plastic waveguide links, several issues still need some attention. Various materials can serve as a dielectric for the waveguide. In this work, only PS, HDPE and PP have been tested and found adequate, but many other materials might be suitable, depending on the requirements. A higher loss per meter may be tolerable if the waveguide is more flexible or can resist higher temperatures. As already hinted in section 5.3, circular or hollow waveguides can also be considered. To support duplex communication, different carrier frequencies can be used, as in [Fuk11], or different polarizations, as in [Dol13]. Contact with a different dielectric causes excessive loss in the waveguide. For example, when the bare waveguide is touched by a finger, losses of more than 20 dB are observed. This can be solved by adding a layer around the waveguide that serves as a dielectric shield. Some experiments are already performed in the lab with simple styrofoam, which acts as a layer of air but prevents direct contact with the waveguide. The loss increases only slightly but unfortunately, due to the larger cross-section, the flexibility decreases so a better solution must be found. The subject of dispersion in plastic waveguide has been touched in subsection 5.2.3, but was not elaborated. The relation between dispersion on the maximum bit rate is difficult to quantify. In the case of the 9 m long waveguide, the maximum data rate for a BER < 10−12 was 2.5 Gbit/s. It is not clear wether this reduction in SNR compared to 9 Gbit/s over 60 cm is due solely to the smaller SNR, caused by the larger loss, or also due to dispersion effects. It would be useful to find a way to measure the dispersion in the waveguide or at least to quantify its effect on the data rate. In the lab, some experiments were performed to explore the possibilities and demonstrate the flexibility of plastic waveguides. It is possible to glue or tape two pieces of waveguide end to end while still ensuring signal propagation. A waveguide splitter was demonstrated by attaching one waveguide to the other. The possibilities of a plastic waveguide as antenna have yet to be explored. Certainly, more interesting properties of plastic waveguides at millimeter wave frequencies will be discovered in the near future.

List of publications

Articles in international journals • M. Tytgat, P. Reynaert,“A 90-GHz receiver in 40-nm CMOS for plastic waveguide links,” submitted to IEEE Transactions on Microwave Theory and Techniques. • M. Tytgat, M. Steyaert, P. Reynaert, “A 186 to 212 GHz downconverter in 90 nm CMOS,” Springer Journal of Infrared, Millimeter, and Terahertz Waves, vol. 33, no. 11, pp. 1085-1103, 2012.

Articles in international conference proceedings • M. Tytgat and P. Reynaert, “A plastic waveguide receiver in 40 nm CMOS with on-chip bondwire antenna,” Proceedings of European Solid-State Circuits Conference, pp. 335-338, Bucharest, Romania, 2013. • V. De Smedt, H. De Clercq, P. Callemeyn, J. Van Rethy, M. Tytgat, J. Verbeeck, B. Puers, M. Steyaert, P. Leroux, G. Gielen, W. Dehaene, “Development of a controller platform for educational projects: A case study,” Proceedings of SEFI, pp. 1-8, Leuven, Belgium, 2013. • M. Tytgat, M. Steyaert, and P. Reynaert, “Time domain model for Costas loop based QPSK receiver,” Ph.D. Research in Microelectronics and Electronics, pp. 313-316, Aachen, Germany, 2012. • M. Tytgat, M. Steyaert, and P. Reynaert, “Generation of Gbit/s modulated millimeter wave signals for measurement,” Proceedings of European Microwave Week, pp. 906-909, Manchester, UK, Oct. 2011. • M. Tytgat, M. Steyaert, and P. Reynaert, “A 200 GHz downconverter in 90 nm CMOS,” Proceedings of European Solid-State Circuits Conference, pp. 239-242, Helsinki, Finland, Sept. 2011.

141

142

LIST OF PUBLICATIONS

Supervised master theses • T. Ma, Y. Zhang, M. Tytgat, P. Reynaert, “A 200 GHz imaging system in 90 nm CMOS,” 2011-2012 • W. Sijbers, M. Tytgat, P. Reynaert, “CMOS geïntegreerde schakelingen voor millimetergolf beeldvorming,” 2010-2011 • S. Kulkarni, M. Tytgat, P. Reynaert, “Design of a Ka-band up-converter for satellite communication,” 2009-2010

Master thesis • M. Tytgat, P. Reynaert, “90 GHz vermogenversterkers in SiGe,” 2008-2009

Bibliography

[Adl73]

R. Adler, “A study of locking phenomena in oscillators”, Proceedings of the IEEE, vol. 61, no. 10, pp. 1380–1385, 1973. pages 119

[Afs08]

B. Afshar, Y. Wang, and A. Niknejad, “A robust 24mW 60GHz receiver in 90nm standard CMOS”, in IEEE International Solid-State Circuits Conference, Digest of Technical Papers (ISSCC), pp. 182 –605, Feb. 2008. pages 16

[Ava12]

Avago Technologies, “Fiber versus copper links: making the right choice for industrial environments”, July 2012. pages 114

[Bor97]

M. Borremans and M. Steyaert, “A 2 V, low power, single-ended 1 GHz CMOS direct upconversion mixer”, in Custom Integrated Circuits Conference, Proceedings of the IEEE, pp. 517–520, 1997. pages 13

[Cho09]

D. Chowdhury, P. Reynaert, and A. Niknejad, “Design considerations for 60 GHz transformer-coupled CMOS power amplifiers”, IEEE Journal of Solid-State Circuits, vol. 44, no. 10, pp. 2733 –2744, Oct. 2009. pages 16, 33

[Cos56]

J. Costas, “Synchronous communications”, Proceedings of the IRE, vol. 44, no. 12, pp. 1713 –1718, Dec. 1956. pages 68, 74

[Cou07]

L. Couch, Digital and Analog Communication Systems, Pearson Prentice Hall, seventh edition, 2007. pages 4, 6, 8, 26, 70, 74, 133

[Cro95]

J. Crols and M. Steyaert, “A 1.5 GHz highly linear CMOS downconversion mixer”, IEEE Journal of Solid-State Circuits, vol. 30, no. 7, pp. 736 –742, July 1995. pages 13, 38, 39, 59

[Def10]

N. Deferm and P. Reynaert, “A 100 GHz transformer-coupled fully differential amplifier in 90 nm CMOS”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 359–362, 2010. pages 121

[Def11a] N. Deferm and P. Reynaert, “A 120GHz 10Gb/s phase-modulating transmitter in 65nm LP CMOS”, in IEEE International Solid-State Circuits Conference, Digest of Technical Papers (ISSCC), pp. 290 –292, Feb. 2011. pages 16, 66, 67, 71, 131 143

144

BIBLIOGRAPHY

[Def11b] N. Deferm, J. Osorio, A. de Graauw, and P. Reynaert, “A 94GHz differential power amplifier in 45nm LP CMOS”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 1–4, 2011. pages 16, 71 [Def13]

N. Deferm, W. Volkaerts, J. F. Osorio, A. de Graauw, M. Steyaert, and P. Reynaert, “A 120GHz fully integrated 10Gb/s wireless transmitter with on-chip antenna in 45nm low power CMOS”, in Proceedings of the IEEE European Solid-State Circuits Conference (ESSCIRC), Sept. 2013. pages 16, 66, 73, 90

[Def14]

N. Deferm, CMOS front ends of millimeter wave wireless communication systems, PhD thesis, KU Leuven, Jan. 2014. pages 37, 73, 78, 85, 90

[Doa05]

C. Doan, S. Emami, A. Niknejad, and R. Brodersen, “Millimeter-wave CMOS design”, IEEE Journal of Solid-State Circuits, vol. 40, no. 1, pp. 144–155, Jan. 2005. pages 17

[Dol13]

N. Dolatsha and A. Arbabian, “Dielectric waveguide with planar multi-mode excitation for high data-rate chip-to-chip interconnects”, in Ultra-Wideband (ICUWB), 2013 IEEE International Conference on, pp. 184–188, 2013. pages 92, 140

[Els49]

W. M. Elsasser, “Attenuation in a dielectric circular rod”, Journal of Applied Physics, vol. 20, no. 12, pp. 1193–1196, 1949. pages 92, 104

[Far10]

Farran Technology, “FTL source modules - datasheet”, http://www.farran.com/index2.php?option=com_ docman&task=doc_view&gid=37&Itemid=154, 2010. pages 59

[Fra13]

Fraunhofer Institute for Applied Solid State Physic, “New world record in wireless data transmission, press release”, http://www.iaf.fraunhofer.de/en/news-media/press_ releases/press-2013-05-16.html, May 2013. pages 112, 138

[Fri44]

H. Friis, “Noise figures of radio receivers”, Proceedings of the IRE, vol. 32, no. 7, pp. 419 – 422, July 1944. pages 23, 40, 41

[Fuk11]

S. Fukuda, Y. Hino, S. Ohashi, T. Takeda, H. Yamagishi, S. Shinke, K. Komori, M. Uno, Y. Akiyama, K. Kawasaki, and A. Hajimiri, “A 12.5+12.5 Gb/s full-duplex plastic waveguide interconnect”, IEEE Journal of Solid-State Circuits, vol. 46, no. 12, pp. 3113 –3125, Dec. 2011. pages 92, 114, 117, 119, 131, 134, 135, 136, 140

[Gar05]

F. M. Gardner, Phaselock techniques, John Wiley and Sons, third edition, July 2005. pages 81, 83

BIBLIOGRAPHY

145

[Gon96]

G. Gonzalez, Microwave Transistor Amplifiers, Prentice Hall, second edition, 1996. pages 20, 21

[Gri08]

D. J. Griffiths, Introduction to electrodynamics, Cummings, third edition, 2008. pages 97, 98

[Gun08]

S. Gunnarsson, N. Wadefalk, J. Svedin, S. Cherednichenko, I. Angelov, H. Zirath, I. Kallfass, and A. Leuther, “A 220 GHz single-chip receiver MMIC with integrated antenna”, Microwave and Wireless Components Letters, IEEE, vol. 18, no. 4, pp. 284 –286, Apr. 2008. pages 16, 65

[Hof03]

A. Hofmann, E. Horster, J. Weinzierl, L.-P. Schmidt, and H. Brand, “Flexible low-loss dielectric waveguides for THz frequencies with transitions to metal waveguides”, in European Microwave Conference (EuMC), vol. 3, pp. 955 – 958, Oct. 2003. pages 92, 104

[Hua11]

S.-J. Huang, Y.-C. Yeh, H. Wang, P.-N. Chen, and J. Lee, “An 87GHz QPSK transceiver with costas-loop carrier recovery in 65nm CMOS”, in IEEE International Solid-State Circuits Conference, Digest of Technical Papers (ISSCC), pp. 168 –170, Feb. 2011. pages 16, 68, 83

[Hub11]

Huber + Suhner, “Datasheet: Coaxial cable: Ez_47_cu_tp”, Jan. 2011. pages 111

[Ina11]

O. Inac, B. Cetinoneri, M. Uzunkol, Y. Atesal, and G. Rebeiz, “Millimeterwave and THz circuits in 45-nm SOI CMOS”, in Compound Semiconductor Integrated Circuit Symposium (CSICS), 2011 IEEE, pp. 1 –4, Oct. 2011. pages 65

[Joh84]

R. C. Johnson and H. Jasik, Antenna Engineering Handbook, McGraw-Hill, second edition, 1984. pages 28, 29, 37

[Joh12]

U. Johannsen, A. Smolders, J. Leiss, and U. Gollor, “Bond-wires: Readily available integrated millimeter-wave antennas”, in European Microwave Conference (EuMC), pp. 197–200, Nov. 2012. pages 124

[Kal11]

I. Kallfass, J. Antes, T. Schneider, F. Kurz, D. Lopez-Diaz, S. Diebold, H. Massler, A. Leuther, and A. Tessmann, “All active MMIC-based wireless communication at 220 GHz”, Terahertz Science and Technology, IEEE Transactions on, vol. 1, no. 2, pp. 477 –487, Nov. 2011. pages 16, 112

Pearson Benjamin

[Kim13] Y. Kim, L. Nan, J. Cong, and M.-C. Chang, “High-speed mm-wave data-link based on hollow plastic cable and CMOS transceiver”, Microwave and Wireless Components Letters, IEEE, vol. 23, no. 12, pp. 674–676, 2013. pages 92, 111, 114, 134, 135, 136

146

BIBLIOGRAPHY

[Kin97]

P. Kinget and M. Steyaert, “A 1-GHz CMOS up-conversion mixer”, IEEE Journal of Solid-State Circuits, vol. 32, no. 3, pp. 370–376, 1997. pages 13

[Kri11]

A. Krishnamoorthy, K. Goossen, W. Jan, X. Zheng, R. Ho, G. Li, R. Rozier, F. Liu, D. Patil, J. Lexau, H. Schwetman, D. Feng, M. Asghari, T. Pinguet, and J. Cunningham, “Progress in low-power switched optical interconnects”, Selected Topics in Quantum Electronics, IEEE Journal of, vol. 17, no. 2, pp. 357–376, 2011. pages 135

[LaR09]

T. LaRocca, J. Y.-C. Liu, and M.-C. F. Chang, “60 GHz CMOS amplifiers using transformer-coupling and artificial dielectric differential transmission lines for compact design”, IEEE Journal of Solid-State Circuits, vol. 44, pp. 1425–1435, May 2009. pages 16

[Las08]

E. Laskin, M. Khanpour, R. Aroca, K. Tang, P. Garcia, and S. Voinigescu, “A 95 GHz receiver with fundamental-frequency VCO and static frequency divider in 65nm digital CMOS”, in IEEE International Solid-State Circuits Conference, Digest of Technical Papers (ISSCC), pp. 180 –605, Feb. 2008. pages 16

[Lee04]

T. H. Lee, The Design of CMOS Radio Frequency Integrated Circuits, Cambridge University Press, second edition, 2004. pages 21

[Lon00]

J. Long, “Monolithic transformers for silicon RF IC design”, IEEE Journal of Solid-State Circuits, vol. 35, no. 9, pp. 1368 –1382, Sept. 2000. pages 33

[Mar69a] E. Marcatili, “Bends in optical dielectric guides”, Bell System Technical Journal, vol. 48, no. 7, pp. 2103–2132, Sept. 1969. pages 92, 105 [Mar69b] E. Marcatili, “Dielectric rectangular waveguide and directional coupler for integrated optics”, Bell System Technical Journal, vol. 48, no. 7, pp. 2071 –2102, Sept. 1969. pages 92, 96, 97 [Max10] Maxim, “Statistical confidence levels for estimating error probability”, http://notes-application.abcelectronique.com/ 003/3-5321.pdf, 2010. pages 25, 131 [Mila]

Millimeter wave products Inc., “Datasheet: 691 series unflanged waveguides”, http://miwv.com/drawings/691/MIWV_Series691. pdf. pages 111

[Milb]

Millitech, “Datasheet: Standard gain conical and pyramial horns”, http://www.millitech.com/pdfs/specsheets/ IS000025-SGH.pdf. pages 112

BIBLIOGRAPHY

[Mit10]

147

T. Mitomo, N. Ono, H. Hoshino, Y. Yoshihara, O. Watanabe, and I. Seto, “A 77 GHz 90 nm CMOS transceiver for FMCW radar applications”, IEEE Journal of Solid-State Circuits, vol. 45, no. 4, pp. 928 –937, Apr. 2010. pages 16

[MIT11] MIT, Microphotonics Center, “Scaling copper 201”, http: //mphotonics.mit.edu/index.php?option=com_ docman&task=doc_download&gid=1219&Itemid=198, Oct. 2011. pages 115 [Miz05]

K. Mizuno, H. Matono, Y. Wagatsuma, H. Warashina, H. Sato, S. Miyanaga, and Y. Yamanaka, “New applications of millimeter-wave incoherent imaging”, in Microwave Symposium Digest, 2005 IEEE MTT-S International, pp. 629–632, June 2005. pages 16

[Moo65] G. E. Moore, “Cramming more components onto integrated circuits”, Electronics, vol. 38, no. 3, pp. 114–117, Apr. 1965. pages 5, 16 [Neu75]

E.-G. Neumann and H. Rudolph, “Radiation from bends in dielectric rod transmission lines”, Microwave Theory and Techniques, IEEE Transactions on, vol. 23, no. 1, pp. 142–149, 1975. pages 105

[Nic08]

S. Nicolson, A. Tomkins, K. Tang, A. Cathelin, D. Belot, and S. Voinigescu, “A 1.2V, 140GHz receiver with on-die antenna in 65nm CMOS”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 229 –232, Apr. 2008. pages 65

[Oje11]

E. Ojefors, B. Heinemann, and U. Pfeiffer, “A 220GHz subharmonic receiver front end in a SiGe HBT technology”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 1 –4, June 2011. pages 16, 65

[OML04] OML inc., “Frequency extension source modules to extend signal generator capability from 50 to 325 GHz”, Microwave Journal, vol. 47, no. 3, Mar. 2004. pages 59, 60 [Osb82]

H. Osborne, “A generalized "polarity-type" Costas loop for tracking MPSK signals”, IEEE Transactions on Communications, vol. 30, no. 10, pp. 2289 – 2296, Oct. 1982. pages 68, 74

[Pfe09]

U. Pfeiffer, E. Ojefors, A. Lisaukas, D. Glaab, and H. Roskos, “A CMOS focal-plane array for heterodyne terahertz imaging”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 433 –436, June 2009. pages 16, 65

[Poz05]

D. M. Pozar, Microwave Engineering, John Wiley & Sons, Inc., third edition, 2005. pages 5, 93

148

BIBLIOGRAPHY

[Rab03]

J. M. Rabaey, A. Chandrakasan, and B. Nikoli´c, Digital Integrated Circuits, Prentice Hall, second edition, 2003. pages 30

[Raz04]

B. Razavi, “A study of injection locking and pulling in oscillators”, IEEE Journal of Solid-State Circuits, vol. 39, no. 9, pp. 1415–1424, 2004. pages 119

[Roh10]

Rohde & Schwarz, “R&S ZVA-Zxx millimeter-wave converters - data sheet”, http://www2.rohde-schwarz.com/file/ZVA-Zxx_ dat_sw_en.pdf, 2010. pages 59

[San06]

W. Sansen, Analog Design Essentials, Springer, 2006. pages 55

[Sin08]

J. Sinsky, A. Konczykowska, A. Adamiecki, F. Jorge, and M. Duelk, “39.4-gb/s duobinary transmission over 24.4 m of coaxial cable using a custom indium phosphide duobinary-to-binary converter integrated circuit”, Microwave Theory and Techniques, IEEE Transactions on, vol. 56, no. 12, pp. 3162–3169, 2008. pages 114

[Ste13]

W. Steyaert and P. Reynaert, “A 0.54 THz signal generator in 40 nm bulk CMOS with 22 GHz tuning range”, in Proceedings of the IEEE European Solid-State Circuits Conference (ESSCIRC), pp. 411–414, Sept 2013. pages 16

[Tak10]

H. Takahashi, T. Kosugi, A. Hirata, K. Murata, and N. Kukutsu, “10Gbit/s quadrature phase-shift-keying modulator and demodulator for 120GHz-band wireless links”, IEEE Transactions on Microwave Theory and Techniques, vol. 58, no. 12, pp. 4072 –4078, Dec. 2010. pages 16, 68

[Tav11]

F. Tavernier, Optical Receivers with Integrated Photodiode in Nanoscale CMOS, PhD thesis, Katholieke Universiteit Leuven, Feb. 2011. pages 94

[Tek09]

Tektronix, “Arbitrary waveform generators - AWG7000B series data sheet”, 2009. pages 61

[Tom09] A. Tomkins, R. Aroca, T. Yamamoto, S. Nicolson, Y. Doi, and S. Voinigescu, “A zero-IF 60 GHz 65 nm CMOS transceiver with direct BPSK modulation demonstrating up to 6 Gb/s data rates over a 2 m wireless link”, IEEE Journal of Solid-State Circuits, vol. 44, no. 8, pp. 2085 –2099, Aug. 2009. pages 16 [Tsi99]

Y. Tsividis, Operation and modeling of the MOS transistor, Oxford University Press, second edition, 1999. pages 29, 31, 46

[Tyt11a] M. Tytgat, M. Steyaert, and P. Reynaert, “A 200GHz downconverter in 90nm CMOS”, in Proceedings of the IEEE European Solid-State Circuits Conference (ESSCIRC), pp. 239 –242, Sept. 2011. pages 38

BIBLIOGRAPHY

149

[Tyt11b] M. Tytgat, M. Steyaert, and P. Reynaert, “Generation of Gbit/s modulated millimeter wave signals for measurement”, in European Microwave Conference (EuMC), pp. 906 –909, Oct. 2011. pages 59 [Tyt12a] M. Tytgat, M. Steyaert, and P. Reynaert, “A 186 to 212 GHz downconverter in 90 nm CMOS”, Springer Journal of Infrared, Millimeter, and Terahertz Waves, vol. 33, no. 11, pp. 1085–1103, 2012. pages 16, 38, 124 [Tyt12b] M. Tytgat, M. Steyaert, and P. Reynaert, “Time Domain Model for Costas Loop Based QPSK Receiver”, in Ph.D. Research in Microelectronics and Electronics (PRIME), 2012 8th Conference on, pp. 313–316, 2012. pages 78 [Tyt13]

M. Tytgat and P. Reynaert, “A plastic waveguide receiver in 40nm CMOS with on-chip bondwire antenna”, in Proceedings of the IEEE European Solid-State Circuits Conference (ESSCIRC), pp. 335–338, 2013. pages 16, 92, 114, 135

[Var07]

M. Varonen, M. Karkkainen, and K. Halonen, “V-band balanced resistive mixer in 65-nm CMOS”, in Solid State Circuits Conference, 2007. ESSCIRC 2007. 33rd European, pp. 360 –363, Sept. 2007. pages 16

[Vir04]

Virginia Diodes, Inc., “New all-solid-state terahertz source”, http://vadiodes.com/VDI/pdf/New%20Terahertz% 20Source%20Nov2004.pdf, 2004. pages 59

[Vol11]

W. Volkaerts, M. Steyaert, and P. Reynaert, “118GHz fundamental VCO with 7.8% tuning range in 65 nm CMOS”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 1 –4, June 2011. pages 16, 66, 124

[Vol13]

W. Volkaerts, M. Steyaert, and P. Reynaert, “A 120GHz quadrature frequency generator with 16.2GHz tuning range in 45nm CMOS”, in IEEE Radio Frequency Integrated Circuits Symposium (RFIC), pp. 207–210, June 2013. pages 16, 66, 73, 90

[Wan06] C.-H. Wang, Y.-H. Cho, C.-S. Lin, H. Wang, C.-H. Chen, D.-C. Niu, J. Yeh, C.-Y. Lee, and J. Chern, “A 60GHz transmitter with integrated antenna in 0.18µm SiGe BiCMOS technology”, IEEE International Solid-State Circuits Conference, Digest of Technical Papers (ISSCC), pp. 659–668, Feb. 2006. pages 16 [Wan08] L. Wang, S. Glisic, J. Borngraeber, W. Winkler, and J. Scheytt, “A singleended fully integrated sige 77/79 ghz receiver for automotive radar”, IEEE Journal of Solid-State Circuits, vol. 43, no. 9, pp. 1897–1908, Sept 2008. pages 16

150

BIBLIOGRAPHY

[Web80] C. Weber and W. Alem, “Demod-remod coherent tracking receiver for QPSK and SQPSK”, IEEE Transactions on Communications, vol. 28, no. 12, pp. 1945 – 1954, Dec. 1980. pages 68, 74 [Wei98]

J. Weinzierl, C. Fluhrer, and H. Brand, “Dielectric waveguides at submillimeter wavelengths”, in Terahertz Electronics Proceedings, 1998. THz Ninety Eight. 1998 IEEE Sixth International Conference on, pp. 166 –169, Sept. 1998. pages 92

[Wik13]

Wikipedia, “Ethernet over twisted pair”, http://en.wikipedia. org/wiki/Ethernet_over_twisted_pair, Aug. 2013. pages 113, 114

[Wik14a] Wikipedia, “Rg-6”, http://en.wikipedia.org/wiki/RG-6, Jan. 2014. pages 113 [Wik14b] Wikipedia, “Category 5 cable”, http://en.wikipedia.org/wiki/ Category_5_cable, Feb. 2014. pages 114 [Yao07]

T. Yao, M. Gordon, K. Tang, K. Yau, M.-T. Yang, P. Schvan, and S. Voinigescu, “Algorithmic design of CMOS LNAs and PAs for 60-GHz radio”, IEEE Journal of Solid-State Circuits, vol. 42, no. 5, pp. 1044 –1057, may 2007. pages 16

[Yeh89]

C. Yeh, F. Shimabukuro, and J. Chu, “Ultralow loss dielectric ribbon waveguide for millimeter/submillimeter waves”, Applied Physics Letters, vol. 54, no. 12, pp. 1183–1185, 1989. pages 104, 111, 134

[Yeh10]

C. Yeh and F. Shimabukuro, The Essence of Dielectric Waveguides, Springer, 2010. pages 92, 96, 97, 98, 99

[Yuj03]

L. Yujiri, M. Shoucri, and P. Moffa, “Passive millimeter wave imaging”, Microwave Magazine, IEEE, vol. 4, no. 3, pp. 39–50, Sept. 2003. pages 16, 29

[Zha13]

D. Zhao and P. Reynaert, “A 60-GHz dual-mode class AB power amplifier in 40-nm CMOS”, IEEE Journal of Solid-State Circuits, vol. 48, no. 10, pp. 2323–2337, Oct. 2013. pages 16

[Zie08]

O. Ziemann, POF Handbook: Optical Short Range Transmission Systems, Springer, 2008. pages 94

FACULTY OF ENGINEERING SCIENCE DEPARTMENT OF ELECTRICAL ENGINEERING ESAT-MICAS Kasteelpark Arenberg 10 3001 Heverlee