millimeter-wave fiber-optic communication system

Research Collection

Doctoral Thesis

Millimeter-wave fiber-optic communication system using InPHEMTs

Author(s): Orzati, Andrea

Publication Date: 2003

Permanent Link: https://doi.org/10.3929/ethz-a-004619835

Rights / License: In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection. For moreinformation please consult the Terms of use.

ETH Library

https://doi.org/10.3929/ethz-a-004619835

http://rightsstatements.org/page/InC-NC/1.0/

https://www.research-collection.ethz.ch

https://www.research-collection.ethz.ch/terms-of-use

MILLIMETER-WAVE FIBER-OPTICCOMMUNICATION SYSTEM USING INP

HEMTS

Andrea Orzati

DISS. ETH No. 15102

DISS. ETH No. 15102

MILLIMETER-WAVE FIBER-OPTICCOMMUNICATION SYSTEM USING INP

HEMTS

A dissertation submitted to theSWISS FEDERAL INSTITUTE OF TECHNOLOGY

ZURICH

for the degree ofDoctor of Technical Sciences

presented byANDREA ORZATI

Laurea in Ing. Elettronica, Universita degli Studi di CagliariBorn July 11, 1973

in Cagliari (Sardinia, Italy)

accepted on the recommendation ofProf. Dr. W. Bachtold, examiner

Prof. Dr. H. Jackel, Prof. M. Vanzi, coexaminers

2003

You see, wire telegraph is a kindof a very, very long cat. You pullhis tail in New York and his headis meowing in Los Angeles. Doyou understand this? And radiooperates exactly the same way:you send signals here, they receivethem there. The only difference isthat there is no cat.

A. Einstein

Acknowledgments

I would like to acknowledge all the people who, in differentways, supported me during my dissertation.In the first place, I would like to thank my “Doktorvater” Prof.Werner Bachtold for giving me the opportunity to work in sucha valid research group. I would also like to express my gratitudeto Prof. Heinz Jackel and Prof. Massimo Vanzi who acceptedto be my coexaminers.I would like to acknowledge Otte J. Homan for his (not only)scientific support, Hanspeter Meier for his patient and indis-pensable work in circuit fabrication, and Hansruedi Benedickterfor revealing me the fine art of high-frequency characterization.My gratitude goes also to Bruno Graf from Avalon Photonicsfor helping us in the dielectric layer deposition, to Iwan Schny-der, for bearing with me the “slings and arrows” of the project,and to Dominique Schreurs for sharing her great experience indevice modeling.A special thank goes to all my friends and colleagues at the in-stitute, for the cheerful working atmosphere they managed tocreate. Thanks to Franck Robin for convincing me to wake upearly to go skiing, climbing, or hiking, to Thomas Brauner forhis patience with my poor sailing skills, to Esteban “Abuelo”Moreno for being always willing to try my latest browniesrecipe, and to Federico Beffa for showing me a delicious al-ternative to the ETH Mensa.Un ringraziamento speciale va anche ai miei genitori, per il loroinesauribile sostegno, alla mia psicologa personale e a tuttele persone che mi sono state vicine in questi anni passati a“dissertare”.

Contents

Acknowledgments v

Abstract xi

Riassunto xiii

1 Introduction 11.1 Motivations 11.2 Project description 41.3 Outline of the work 4

2 The MFOCS System 72.1 System Design 7

2.1.1 Important Terminology 72.1.2 Guidelines for the Design of mm-wave

fiber-optic communication system 82.1.3 Different System Concepts 112.1.4 Proposed Approach 15

2.2 Hub Design 172.2.1 Analysis of the Fiber-Optic Data Link

172.2.2 Analysis of the Radio Link 192.2.3 Hub Architecture 23

2.3 Conclusions 25

3 Process Technology 273.1 The High Electron Mobility Transistor 27

3.1.1 From the two-dimensional electron gasto InP HEMTs 27

viii Contents

3.1.2 Physical principles of heterostructures28

3.1.3 HEMT Material Systems 293.1.4 Device structure 31

3.2 Process Description 323.3 Components 35

3.3.1 InP HEMTs 353.3.2 Transmission Lines 373.3.3 Line Discontinuities 403.3.4 Thin Film Resistors 443.3.5 MIM Capacitors 443.3.6 Spiral Inductors 44

4 InP HEMT Large Signal Model 474.1 Large-Signal Modeling of High-Frequency

FETs 484.2 Small-Signal and Large-Signal Equivalent

Circuit 504.3 Model Extraction 554.4 Model Validation 58

4.4.1 Small-signal verification 584.4.2 Large-signal verification 614.4.3 Large-Signal Validation through MMIC

Design 644.5 Limitations of Look-up Table Based Large-

Signal Models 684.6 Conclusions 69

5 MFOCS Circuits 715.1 0-20 GHz Traveling Wave Amplifier 71

5.1.1 Introduction 715.1.2 Circuit Design 725.1.3 Measurement Results 75

5.2 16 GHz Up-converting Image-Rejection Re-sistive Mixer 775.2.1 Introduction 77

Contents ix

5.2.2 Circuit Design 775.2.3 Measurement Results 79

5.3 16-48 GHz Active Frequency Triplers 835.3.1 Introduction 835.3.2 Theoretical Analysis 835.3.3 Circuit Design 895.3.4 Measurement Results 92

5.4 V-band Up-Converting Active Mixer 975.4.1 Introduction 975.4.2 Circuit Design 975.4.3 Measurement Results 99

6 Flip-Chip Bonding 1036.1 General Aspects 1046.2 Why Flip-Chip 1046.3 Process description 1066.4 Optimization for V-band applications 1076.5 Application to InP MMICs 1136.6 Encountered Problems 1166.7 Conclusions 116

7 Summary, Conclusions, and Outlook 1197.1 System Design 1197.2 InP HEMT Large-Signal Modeling 1207.3 Circuit Design and Fabrication 1207.4 Flip-Chip optimization for V-band Appli-

cations 1217.5 Conclusions and Future Work 122

Bibliography 125

Abstract

Goal of the work presented in this dissertation is to con-tribute to the technology, the design, and the characteriza-tion of a 60 GHz transceiver for a fiber-optic millimeter-wavewireless LAN. This was done in the frame of the MFOCS(Millimeter-wave Fiber-Optic Communication System) project,funded by the the Swiss Federal Institute of Technology.Through an analysis of the characteristics of both the opti-cal link and the radio link, a feasible system architecture andrealistic hub specifications can be derived. Remote generationof the mm-wave local oscillator through frequency multiplica-tion in the hub is proposed as the best way to generate a stablehigh-frequency LO signal and contemporaneously overcome theproblems due to chromatic dispersion in the fiber. A two-stepsignal up-conversion in the hub is suggested instead of a directup-conversion in order to relax the mixer specifications in termsof LO- and image-rejection and allow the frequency conversioncircuits to be fabricated on the InP HEMT process of the Mi-crowave Electronics Laboratory.In order to design and fabricate integrated circuits that meetthe required specifications, precise models for both passive andactive components are indispensable. Transmission line elementswere modeled by means of an electromagnetic simulator. ForInP HEMTs, a look-up table-based large-signal model was de-veloped which takes impact ionization into account. S-parametermeasurements demonstrate that the linear performance of themodel is excellent up to 110 GHz. The non-linear performanceis validated in the impact ionization region as well as up to 64GHz by non-linear measurements at different bias points.Using the developed model, different circuits were designed and

xii Abstract

fabricated. They include a 0 − 20 GHz optical receiver ampli-fier, a K-band up-converting image-rejection resistive mixer,a V-band up-converting active mixer, and a 16 − 48 GHz fre-quency tripler. The measured performances of the fabricatedcircuits are in line with the required system specifications.In order to assemble the system, a simple low-cost gold-ballflip-chip bonding process was optimized for V-band applica-tions. For the optimized CPW-CPW flip-chip transition, returnlosses lower than −20 dB up to 80 GHz were measured. Theapplication of the optimized bonding technique to microwaveamplifiers was also demonstrated.

Riassunto

Il lavoro di ricerca confluito nella presente dissertazione vuoleessere un contributo allo sviluppo della tecnologia, delle tec-niche di progettazione e delle tecniche di caratterizzazione diun ricetrasmettitore per reti locali che utilizzi sia fibre otticheche segnali radio ad onde millimetriche. Per la realizzazione delsistema si e proposto l’uso di circuiti integrati in fosfuro d’indiobasati su transistor ad alta mobilita elettronica (HEMT), re-alizzati usando il processo sviluppato dal Laboratorio di AltaFrequenza (IFH). La ricerca e stata finanziata dal PolitecnicoFederale di Zurigo, nell’ambito del progetto MFOCS (Millimeter-wave Fiber-Optic Communication System).L’architettura e le specifiche del sistema sono state derivateda attente analisi del collegamento su fibra ottica e del canaleradio. Nel ricetrasmettitore, la generazione del segnale dell’o-scillatore locale a onde millimetriche avviene tramite la molti-plicazione di un segnale di riferimento a bassa frequenza rice-vuto attraverso la fibra ottica. In questo modo vengono con-temporaneamente risolti i problemi della dispersione cromaticanella fibra ottica e della generazione di un segnale di riferi-mento ad onde millimetriche con i giusti requisiti di stabilita.Nel ricetrasmettitore, la conversione di frequenza viene effet-tuata in due stadi; questa soluzione ha l’effetto di rilassare lespecifiche dei circuiti che operano la miscelazione del segnale,sia dal punto di vista della reiezione delle componenti spuriedovute sia all’oscillatore locale che alla frequenza immagine.Per poter progettare e, quindi, realizzare circuiti integrati cherientrino nelle specifiche richieste dal sistema, e indispensabiledisporre di modelli circuitali precisi, sia per i componenti pas-sivi che per gli HEMT. Le discontinuita delle linee di trasmis-

xiv Riassunto

sione sono state modellate con successo usando un simulatoreelettromagnetico. Per gli HEMT, e stato realizzato un modelloa largo segnale basato su look-up table che tiene conto deglieffetti della ionizzazione da impatto. Misure dei parametri didispersione hanno dimostrato come la capacita di predizionedel comportamento lineare del transistor sia ottima fino a 110GHz. Il comportamento a largo segnale del modello e stato ve-rificato, per diversi punti di lavoro, sia a basse frequenze, dovee osservabile l’effetto della ionizzazione da impatto, sia a fre-quenze delle onde millimetriche (64 GHz).Con l’ausilio dei modelli sviluppati, i diversi componenti delsistema sono stati progettati e realizzati nella camera biancadell’istituto. Questi circuiti sono: un amplificatore distribuitoper il foto ricevitore con una banda passate da 200 MHz a 20GHz, un up-converting mixer resistivo in banda K a reiezionedella frequenza immagine, un up-converting mixer attivo inbanda V ed un moltiplicatore di frequenza da 16 a 48 GHz. Lecaratteristiche di tutti i circuiti misurati rientrano nelle speci-fiche dettate dal sistema.Allo scopo di facilitare l’assemblaggio del sistema, e stato svilup-pato e ottimizzato per applicazioni in banda V un processo dimontaggio a chip rovesciato (flip-chip). In questo modo e statopossibile ottenere transizioni tra il chip e la scheda di montag-gio con coefficienti di riflessione misurati inferiori ai −20 dBper frequenze fino ad 80 GHz. Questa tecnica e stata usata consuccesso per il montaggio su scheda di amplificatori integrati amicroonde.

1

Introduction

1.1 Motivations

Wireless systems represent a significant part of today telecom-munication business. This is mainly due to the extraordinarytechnical and commercial success experienced by mobile phonesin the last ten years. These days, thanks to the increasing popu-larity of portable personal computers and data entry terminals,academic and corporate interest has gradually shifted to wire-less data systems and wireless local area networks (WLANs).Second generation WLANs offering data rates between 10 and25 Mb/s are in a phase of advanced development (wirelessATM), if not already on the market (wireless Ethernet). Nev-ertheless, there is still a multitude of multimedia applicationsrequiring wireless transmission over a short range which can notbe satisfied by existing communication systems [1, 2]. Some ofthe most significant short-range wireless applications are sum-marized in Table 1.1.The frequency band around 60 GHz has been indicated as

the most convenient for short-distance wireless communica-tions [1, 3]. As shown in Fig. 1.1, in a 8 GHz band centeredaround 60 GHz an attenuation peak (10 − 15 dB/km) due toatmospheric oxygen limits communication distance to less thanone kilometer. As a result, this portion of the spectrum allowsheavy frequency reuse, particularly in an indoor environment,where the signal propagation is de facto heavily conditioned bythe geometry of the building [4]. These are all reasons why inEurope, in the United States, as well as in Japan, huge band-

2 Introduction

Table 1.1 Some of the most significant short-range wireless applications.

Application Capacity per user (Mb/s)

Wireless LAN bridge 100 − 1000Virtual reality allowing free bodymovements

450

Wireless IEEE 1394 100, 200, 300Wireless High-resolution recordingcamera

150 − 270

Trading terminal having multiplevideo channels that can be viewedsimultaneously

50 − 100

Wireless News tablet 50 − 100Internet download of lengthy files 10 − 100High-quality video-conference 10 − 100Ad hoc communication, i.e., directcommunication between notebooks,between notebook and nearbyprinter, etc.

0.1 − 100

widths have been allocated around the 60 GHz band, whichcan be therefore proficiently exploited for very high data ratestransmissions [5, 6].Next generation WLANs will be found in offices as well as athome and will consist of various pico-cells, each one covering asingle room. In a typical installation of a future WLAN, therewill be a large number of fixed access points connected to somebase stations through a fixed cable infrastructure. Optical fibersare the natural candidate to link access points with base sta-tions because of their excellent characteristics in terms of band-width and propagation losses [7]. It is, therefore, not contro-versial to expect that fiber-optic-fed millimeter-wave wirelesssystems will be a central element of future networking infras-tructures. The data will be distributed through a building bymeans of optical fibers and received by hubs responsible for theoptical-electrical conversion and for the signal transmission, atmillimeter-wave frequencies, to the final users. An example of

Motivations 3

101

102

103

10−2

10−1

100

101

102

103

Frequency (GHz)

Atm

osph

eric

Atte

nuat

ion

(dB

/km

)

O2

60 GHz

Fig. 1.1 Atmospheric attenuation of microwaves as a function of frequency.

a possible future office environment is illustrated in Fig. 1.2.High electron mobility transistors (HEMT) based on the InP

material system have have proved to be of high interest formillimeter-wave applications. In the V-band, compared to GaAspseudomorphic HEMTs (PHEMTs), they have shown highergain, lower noise figure, and, in spite of their lower breakdownvoltage, better power added efficiency [8–10]. For these rea-

Fig. 1.2 A 60 GHz wireless LAN system in an office environment.

4 Introduction

sons, in the last fifteen years this field of research has been theobject of constant attention at the Swiss Federal Institute ofTechnology in Zurich (ETHZ). Significant advantages are ex-pected from employing InP HEMTs technology in millimeter-wave wireless communication systems.Goal of the project described in this work was to contribute tothe technology, the design, and the characterization of a 60 GHztransceiver for a fiber-optic millimeter-wave wireless LAN. Thiswas done without the pretense of delivering a commercially at-tractive product, but with the intention of contributing, byshowing its advantages, to the establishment of a mature InPHEMT circuit technology at ETHZ. The project has been car-ried out using the in-house InP HEMT process of the Labo-ratory for Electromagnetic Fields and Microwave Electronics(IFH). The following chapters will present the author’s effortsand achievements during the last four years.

1.2 Project description

The research work upon which this dissertation is based wascarried out within the MFOCS (Millimeter-wave Fiber-OpticCommunication System) project, directly funded by the theSwiss Federal Institute of Technology. Goal of the project wasto contribute in the technology, design, and characterization ofan integrated transceiver for a fiber-optic millimeter-wave wire-less LAN. The project was undertaken in collaboration with theElectronics Laboratory (IfE) of the Swiss Federal Institute ofTechnology in Zurich, which runs an InP HBT process, andwas aimed to support InP technology at ETHZ. IFH was re-sponsible for the transceiver design which used InP HEMTs,while IfE attended to the generation and transmission of theoptical signal through the fiber, which was implemented on anInP HBT process.

1.3 Outline of the work

The present dissertation is structured as follows. In Chapter2 the system design constrains are described and the chosen

Outline of the work 5

system architecture is presented and motivated. Chapter 3describes the in-house InP HEMT process employed for theMMICs’ fabrication, its features, and its limitations. Chapter4 presents the InP HEMT large-signal model developedduring the project and used for circuit design; the model takesimpact ionization into account and was validated up to the V-band by linear and non-linear measurements. All the MMICs

that have been designed, fabricated, and characterized are pre-sented in Chapter 5; their design issues are discussed togetherwith their measured performance. Chapter 6, finally, presents aflip-chip mounting technique optimized for V-band appli-cations and its application to millimeter-wave coplanar circuits.

2

The MFOCS System

The present chapter describes the design flow that, startingfrom unrefined considerations, led to the development of thesystem architecture and, from there, to concrete circuit speci-fications. This chapter is structured as follows. In the first sec-tion, the architecture of the complete system is presented.After defining some key terms often used in the rest of the chap-ter, the practical aspects limiting the designer freedom in thechoice of the system architecture are discussed. Finally, after areview of the most significant approaches found in the litera-ture, the proposed system is presented.In the second part of the chapter the attention is focused onthe transceiver, whose development was the actual goal of thework. In the description of how the transceiver specificationsare derived, also typical engineering issues are discussed: theanalysis of the fiber-optic data link and the radiation constraintgiven by the current legislation. At the end of the chapter, thetransceiver architecture is presented and a description ofthe needed circuits is given.

2.1 System Design

2.1.1 Important Terminology

At the beginning of this section it is needful to univocally de-fine some terms that will be often used later in the chapter.The section, which has not the pretension to be systematic orexhaustive, contains a list of important terms, associated with

8 The MFOCS System

their meaning in the context of this work. This is necessary be-cause of the lack of harmonization in the terminology found inthe current literature and will provide an easier understandingof the following sections.

• Base station. It is the part of the system responsible for dis-tributing the information coming from a data source (e.g.a server) to a number of remote hubs and vice versa. Thebase station is wired to the different hubs through , e.g.,optical fibers (Fig. 2.1).

• Hub. It is the interface between the final user (e.g. a mo-bile computer) and the communication system (Fig. 2.1). Itprovides a bidirectional link with both the base station andthe final users. In a communication system there are severalhubs connected with the same base station. As sometimesfound in literature, this part of the system can also be calledtransceiver or transponder or access point.

• Mobile receiver. It is the final user of the WLAN (Fig. 2.1).

• Picocell. It is the portion of space served by a single hub(Fig. 2.1). In a 60 GHz WLAN the picocell usually coincideswith a single room, due to the propagation characteristicsof millimeter-waves at these frequencies.

• Down-link. It is the unidirectional radio link (in case of aWLAN) that connects the hub with the final user (Fig.2.1).

• Up-link. Analogous to the down-link, but directed from thefinal user to the hub (Fig. 2.1).

2.1.2 Guidelines for the Design of mm-wave fiber-

optic communication system

In the design of a millimeter-wave fiber-optic communicationsystem there are various practical issues that have to be takeninto account and that constrain the choice of the system de-signer. These factors are of disparate nature and will be illus-trated in the following sections.

System Design 9

Fig. 2.1 Schematic representation of the different components of a WLAN.

Current Regulation on Frequency Allocation

Frequency allocation in the mm-wave region changes dependingon the country. While in Japan the Multimedia Mobile AccessCommunication (MMAC) committee is still deciding how toexploit the advantages of the 60 GHz band, in Europe andin the United States the European Commission for Posts andTelecommunications (CEPT) and the Federal CommunicationsCommission (FCC), respectively, already proposed an ad hocregulation on this matter [5,6]. In Europe, the 62−63 GHz andthe 65− 66 GHz bands are dedicated to WLAN down-link andup-link, respectively. In the United States, the 59 − 64 GHzfrequency band was set aside for general unlicensed applica-tions. This is the largest contiguous block of radio spectrumever allocated. In this work, the recommendations of the Euro-pean Commission for Posts and Telecommunications are takenas reference.

Limitations due to Chromatic Dispersion

The term chromatic dispersion indicates the wavelength de-pendence of the optical-signal propagation velocity on the fibermaterial. Since every optical signal has a finite spectral width,

10 The MFOCS System

this dispersion results in a spreading of the signal spectrum.In intensity-modulated direct-conversion links operating above20 GHz, chromatic dispersion significantly limits the transmis-sion distance by introducing a carrier to noise penalty in thetransmitted signal [11]. Due to chromatic dispersion, an intensity-modulated 60 GHz signal would be completely extinct afterone kilometer. In [12] the transmission penalty due to chro-matic dispersion in a single-mode fiber (SMF) is calculated asa function of the signal frequency and of the fiber length. It isshown that for a low-penalty transmission over 10 km SMF themodulation frequency can not be higher then 20 GHz. Numer-ous methods have been studied to overcome the effects of chro-matic dispersion. Smith and Novak, using an approach based onoptical single-sideband modulation techniques [13, 14] demon-strated a 51.8 Mb/s digital transmission at 12 GHz over 80 kmof single-mode fiber. Other methods based on optical hetero-dyning [11,15] were also successful.

Generation of a Stable High-Frequency Local Oscillator Signal

A crucial point in the design and implementation of mm-wavefiber-optic communication systems is the generation of millimeter-waves in the hub, since their direct transmission from the basestation over the fiber encounters chromatic dispersion. A pos-sible way to skirt the issue would be to have a high-frequencylocal-oscillator (LO) in the hub and use it to mix the datasignal up to 60 GHz. This solution is, for technological rea-sons, not as practicable as it seems in a first moment. In fact,millimeter-wave oscillators with the reproducibility and stabil-ity characteristics required by the application are not feasible.Possible alternatives are the use of frequency multiplicationchain together with a stable low-frequency LO source or ofvoltage controlled oscillators with a control loop. Both solu-tions drastically increase the hub complexity and, therefore, itscost and can not considered valid. It is clear that the generationof a stable high-frequency local oscillator can not be performedin the hub but other solutions have to be found for it.

System Design 11

2.1.3 Different System Concepts

In this section, an overview of different approaches to the designof millimeter-wave fiber-optic system is presented. The atten-tion is focused on the down-link implementation. For almost allthe described systems, in fact, the up-link and the down-linkhave analogous implementations, and a description of both ofthem would just cause redundancy. In case up- and down-linkshow significant differences, it will be brought to the readersattention.

Millimeter-waves over the fiber

This approach was proposed by Imai and Kawamura in 1995[16,17]. The idea is to modulate the data signal directly in thebase station using a 43.75 GHz LO and to send the modulatedsignal over the fiber. In the hub, the optical signal is received,amplified, and transmitted via antenna to the mobile receiverwithout the need for any frequency conversion. This methodhas the advantage of keeping the hub architecture at minimumcomplexity, but has also some significant disadvantages. In thefirst place, the fiber length, i.e. the distance between the basestation and the hub is strongly limited by chromatic dispersion,which is already very strong at 43.75 GHz. Additionaly, veryhigh frequency photodetectors, lasers, and laser-modulators areneeded, not to mention the fact that it would be very difficultto extend this approach to the 60 GHz band. Nevertheless,the mentioned authors were able to successfully demonstratea 118 Mb/s digital transmission over 50 m optical fiber and4 m wireless link. A schematic representation of the proposedsystem is depicted in Fig. 2.2(a).

Wavelength Division

This method was proposed in 1997 by Noel and is describedin [18]. A broad-band (120 Mbyte/s QPSK data and 20 digitalTV channels) 60 GHz transmission over a 13 Km optical fiberand a 5 m radio link was demonstrated. The system works asfollows: a 59 GHz LO is generated in the base station using a

12 The MFOCS System

master/slave distributed-feedback laser configuration and sentover the fiber. This signal has a very high spectral purity whichmakes it almost insensitive to chromatic dispersion. The datasignal is in the 0.8 − 2 GHz band and is transmitted using adistributed-feedback laser, separated from the LO in the wave-length domain. An electroabsorption modulator (EAM) is usedin the hub to receive (and transmit) the data signal over thefiber, while the carrier is detected in the hub by means of aphoto diode. The LO signal is used to up-convert the data sig-nal to the 60 GHz band. The generated millimeter-wave signalis then sent to a high-gain antenna.This approach overcomes the effects of chromatic dispersionin the fiber by using a high-purity LO source and by separat-ing data signal and carrier in the wavelength domain. On theother hand, high-cost components such as an electroabsorptionmodulator, a wavelength division multiplexer, and a 60 GHzphotoreceiver have to be used in the hub in order to obtainthese, indeed, excellent results. This approach is schematicallyrepresented in Fig. 2.2(b).

Optical Heterodyning

A demonstration of a system based on the optical heterodyn-ing principle was given by Braun in 1998 and is presentedin [19,20]. The base station generates two optical signals with afrequency spacing equal to the frequency of desired millimeter-wave signal. If one of these laser is amplitude modulated withthe data signal and if coherent detection in the hub is guar-anteed, then the beating of the two optical signal generates amillimeter-wave carrier amplitude-modulated by the data sig-nal. There are different ways of generating the two beatingoptical signals, both with single optical sources (modulationsideband techniques, mode locked laser, dual-mode laser, FM-modulated laser in conjuntion with fiber dispersion) and withmultiple optical sources. This method guarantees a simple andeffective down-link architecture and is very well suited for appli-cations where only unidirectional transmission is required, but

System Design 13

Fig. 2.2 Schematic representation of different system concepts.

14 The MFOCS System

still presents open issues when it comes to bidirectional trans-mission. It is possible using optical sideband injection lockingto generate a high-frequency LO signal in the hub, but it stillremains complicated to make an effective use of it. A simplifiedversion of the system is given in Fig. 2.2(c).

Subharmonic Up-conversion

This approach was implemented by Smith and Novak in 1998and its detailed description can be found in [21]. They demon-strated a full-duplex transmission over over 40 km of standardsingle mode fiber and 5 m radio link. A schematic represen-tation of the system can be found in Fig 2.2(d). In the basestation, a data signal in the 2 − 3 GHz band is combined witha 17 GHz LO and the resulting signal is then transmitted overthe fiber. In the hub the signal is detected and the LO is sepa-rated from the data by means of a diplexer. Then, the LO signalis used in a sub-harmonically pumped image-rejection mixer toconvert the data up to the 40 GHz band. This approach solvesthe problem of chromatic dispersion by limiting the frequencyof the signal sent over the fiber to 17 GHz and has the ad-vantage of presenting a simple hub architecture. However, thismethod can arduously be extended to the 60 GHz band just byusing sub-harmonically pumped mixers.

Frequency Multiplication in the Hub

In 1998 a novel system design and transmission experimentswere presented by Kojucharow [22,23], where the digital trans-mission of 50 Mb/s over a 4.5 m radio link was successfullydemonstrated. A schematic representation of the proposed ap-proach is depicted in Fig. 2.2(d). The base station design isanalogous to the one proposed by Smith in [8] and describedin the previous section. A data signal modulated around 2.223GHz is combined with a 4.8 GHz local oscillator by means ofan EAM. The resulting signal is then sent over the fiber anddetected by a photo diode at the hub. After reception the LOcomponent of the signal is separated from the data using a

System Design 15

microstrip diplexer. The LO is them multiplied in frequencyby a factor of twelve and used to up-convert the data signalto the V-band. This approach has several advantages; effectsof chromatic dispersion are minimized by limiting to 4.8 GHzthe highest frequency transmitted over the optical fiber, thehub architecture is kept quite simple, and no high-performanceoptical component is required.

2.1.4 Proposed Approach

A schematic representation of the proposed system is show inFig. 2.3. This design is an extension of the approach suggestedby Kojucharow and described in [22, 23]. Some modificationswere introduced in order to take the current European reg-ulations on frequency allocation into account and to furthersimplify the hub architecture.In the base station, a 155 Mb/s NRZ data signal modulates thephase of a 1.5 GHz carrier. The modulated signal is then com-bined with a 16 GHz local oscillator, and the resulting signalis given as input to a laser driver and sent through the opticalfiber. Since the maximum modulation frequency of the opti-cal signal is 16 GHz, no significant penalty due to chromaticdispersion is expected, if the fiber length is kept below 10 km.Using 16 GHz as reference frequency in the base station insteadof 4.8 GHz, facilitates the generation of a high-frequency localoscillator in the hub. Since no frequency multiplication chainis needed, the result is a simpler architecture. Furthermore, a10 km transmission distance over the fiber is more than whatcan be expected for this kind of application.In the hub, the optical signal is detected, amplified, and sepa-rated by a diplexer into its data and LO components. A powersplitter divides the 16 GHz LO signal in two; a part is used tomix the data up to 14.5 GHz, the other is converted by a fre-quency tripler into a 48 GHz reference signal, which is then usedto up-convert the data signal up to 62.5 GHz, i.e. in the middleof the frequency band allocated for the down-link. In contrastwith the approaches presented in the previous sections, the up-

16 The MFOCS System

Fig. 2.3 Schematic representation the proposed system.

conversion of the data signal is performed in two successivesteps. The reason of this choice lies in the extreme difficultyof designing a monolithically integrated V-band mixer able todirectly up-convert the 1.5 GHz data signal to 62.5 GHz and en-sure at the same time an effective suppression of the 64 GHz LOtogether with a good image rejection (65.5 GHz). The qualityfactors required by the filtering networks, unfortunately, do nobelong to this world. Additionally, the two-step up-conversionscheme only requires an additional mixer if compared to directup-conversion. The generated 16 GHz and 48 GHz carriers canbe also employed to down-convert the up-link signal.

Hub Design 17

2.2 Hub Design

In the previous sections, an approach to the development ofa millimeter-wave fiber-optic communication system was pro-posed. In this second part of the chapter, the attention is fo-cused on the hub, which was the main topic of the work leadingto this dissertation. The hub architecture was already intro-duced in the previous sections. Nevertheless, to derive correctspecifications for the transceiver, a deeper insight of how thispart is interfaced with the rest of the system is necessary. Thehub is connected to the base station by an optical-fiber link andcommunicates with the mobile receiver through a radio chan-nel. To lay down a proper set of specifications for the hub, ananalysis of these two links is, therefore, essential.

2.2.1 Analysis of the Fiber-Optic Data Link

In this section, a simple analysis of the optical data link is pre-sented, which can be used to derive coarse specifications for theoptical power transmitted from the base station to the hub asa function of the receiver noise figure (NF). This is done underthe assumption that the data signal is phase modulated usinga antipodal phase shift keying (2-PSK) and has a bandwidth Bof 1 GHz. The transmission occurs over a 10 km single-modefiber with an attenuation of 0.25 db/km, for the transmittinglaser a relative intensity noise (RIN)of −137 dB/Hz was as-sumed, and the responsivity ηPD of the photo diode was setto 0.4 A/W. These values are extracted from the data sheetsof commercially available components. To simplify the calcula-tions, the optical modulation index is considered to be one andthe laser non-linearities, as well as the finite extinction ratioand the dark current are not considered.For an antipodal 2-PSK modulation the bit error probability(BEP) is given by equation 2.1 [24]:

BEP =1

2erfc

(

√

SNR

2

)

. (2.1)

18 The MFOCS System

Fig. 2.4 Noise equivalent circuit of an optical receiver.

According to the International Telecommunication Union - Telecom-munication Standardization Sector (ITU-T) recommendations( [25]), a good transmission quality requires a BEP smaller than10−10, which, from equation 2.1, leads to a SNR greater than13 dB. Figure 2.4 shows the noise equivalent circuit of a opticalreceiver [26]. The light power Popt coupled into the photo diodegenerates an electrical current ηPDPopt at the receiver inputwhich leads to a quantum shot noise current:

〈I2sh〉 = 2qηPDPoptB, (2.2)

where q is the elementary charge. Another noise source derivesfrom the relative intensity noise of the transmission laser:

I2RIN = RIN(ηPDPopt)

2B. (2.3)

The noise contribution of the amplifier can be calculated fromits noise figure (NF) and from its equivalent input resistanceRamlpi. The amplifier noise current can be expressed as:

I2n =

4kTB

Rampli

·(

NF − 1

)

, (2.4)

where k is the Boltzmann constant (1.38 · 10−23 J/K) and Tthe temperature in Kelvin. The total noise current becomes:

In,tot =√

〈I2sh〉 + 〈I2

R〉 + I2RIN + I2

n, (2.5)

and the SNR can be expressed as:

Hub Design 19

0 2 4 6 8 1010

1315

20

25

30

35

40

45

50

Amplifier Noise Figure (dB)

SN

R (

dB)

Pt = 0.1 mW

Pt = 1 mW

Pt = 10 mW

Fig. 2.5 Noise equivalent circuit of an optical receiver.

SNR = 20log

(

PoptηPD

In,tot

√B

)

. (2.6)

Figure 2.5 shows the SNR as a function of the amplifier noisefigure for a transmission power ranging from 0.1 to 10 mW. Itcan be deduced that the requirements on the amplifier noisefigure are quite relaxed for this kind of application.

2.2.2 Analysis of the Radio Link

In this section, the different aspects of the radio link betweenthe hub and the mobile receiver are investigated. This is nec-essary not only to determine the minimum transmitted powerwhich is required for a given bit-error-rate, but also to set itsupper limit according to the existent regulations on electro-magnetic emissions.

Aspects Related to Current Regulations on Electromagnetic

Emissions

The Swiss Federal Council, in an ordinance relating to protec-tion from non-ionising radiation [27], set 1 mW/cm2 as maxi-mum allowed power density at 60 GHz. In Fig. 2.6 power den-

20 The MFOCS System

0 5 10 15 20 25 300

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

Transmitted Power (dBm)

Pow

er D

ensi

ty (

mW

/cm

2) 25 dBi

20 dBi

15 dBi

10 dBi

Fig. 2.6 Power Density at 0.5 m distance from the transmitter for four different

antenna gains.

0 5 10 15 20 25 300

5

9

10

15

Transmitted Power (dBm)

Pow

er D

ensi

ty (µ

W/c

m2 )

25 dBi

20 dBi

15 dBi

10 dBi

Fig. 2.7 Power Density at 3 m distance from the transmitter for four different

antenna gains.

Hub Design 21

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−102

−100

−98

−96

−94

−92

−90

Signal Bandwidth (GHz)

Pm

in (

dBm

)

Fig. 2.8 Minimum received power as a function of the signal bandwidth for a

2-PSK modulation.

sity at 0.5 m distance from the transmitter is plotted as a func-tion of transmitted power for four different antenna gains. Inorder to trespass the limits set in [27] more than 20 dBm trans-mitted power in combination with a 25 dBi antenna is needed.

Rules set by the Federal Commission for Communication

In the United States, the Federal Commission for Communica-tion (FCC) allows 10 W of equivalent isotropic radiated power(EIRP) in the 60 GHz frequency band [6]. This correspondsto a maximum power density of 9 µW/cm2 at 3 m distancefrom the transmitter. Given that similar rules are going to beadopted in Europe as well, this sets another upper limit for thetransmitted power. Figure 2.7 shows the power density at 3 mfrom the antenna for four different antenna gains. It can benoticed that these recommendations are more restrictive thanthe ones in [27].

22 The MFOCS System

1 2 3 4 5 6 7 8 9 10−90

−85

−80

−75

−70

−65

−60

−55

−50

−45

Transmission Distance (m)

PR

(dB

m)

Pt=5 dBm

Pt=10 dBm

Pt=15 dBm

Fig. 2.9 Received power as a function of transmission distance.

Link Budget

In order to determine the minimum transmitted power whichis required for a satisfactory link performance, a two-way linkequation was adopted, which has some modifications in orderto take additional losses at millimeter-wave frequencies intoaccount [28, 29]. The minimum allowed received power can beexpressed as:

Pmin = 10log(kTB) + NFrec + SNRmin, (2.7)

where k is the Boltzmann constant, T the room temperature inKelvin, B the signal bandwidth, NFrec the noise figure of themobile receiver, and SNRmin is the minimum allowed SNR. Inequation 2.7, 10log(kTB) represents the gaussian noise addedby the radio channel. In Fig. 2.8 the minimum received powerfor a NFrec of 10 dB and a SNRmin of 13 dBm (2-PSK with10−10 BER) is shown as a function of the signal bandwidth.According to [28] the received power can be calculated as:

Prec =PtGtGrλ

2Fm

16π2Ra. (2.8)

Hub Design 23

In the 2.8, Pt is the transmitted power, Gt and Gr the antennagains of the transmitter and of the receiver, respectively, λ isthe operating wavelength (0.0045 m), Fm is a link margin in-troduced to compensate non-ideal effects, R is the transmissiondistance, and a is a loss exponent that can be experimentallydetermined. Assuming a total antenna gain (GtGr) of 20 dBi, alink margin of −15 dB, and a loss exponent of 3, we calculatedthe curves of Fig. 2.9, where the received power is shown asa function of the transmission distance for three different val-ues of the transmitted power. It can be observed that 5 dBmtransmitted power are enough to ensure the transmission of a2-PSK modulated signal with a bandwidth of 1 GHz.The hub transmission power was chosen to be 10 dBm, whichlies on the middle between the minimum and the maximumpermitted power and represents a feasible goal with the avail-able technology.

2.2.3 Hub Architecture

The link equations described in the previous sections were usedto develop the fuzzy hub description given in Section 2.1.4 intoa precise and feasible design. Figure 2.10 presents the final de-sign of the hub. The marked circuits were developed during theproject and will be presented in this dissertation. The circuits

Fig. 2.10 Schematic representation of the proposed hub architecture.

24 The MFOCS System

needed for the realization of the hub are:

• 0-20 GHz Traveling Wave Amplifier (TWA). It amplifiesthe electrical signal generated by the photo diode. Its spec-ifications are quite relaxed both in terms of noise figure andof gain.

• Active Diplexer. This component takes the signal comingfrom the TWA and separates the frequency components inthe 0−2 GHz band from the 16 GHz LO. It is called activebecause the LO signal is also amplified.

• K-Band Resistive Mixer. It mixes the data signal from the1 − 2 GHz band up to the 14 − 15 GHz band. It has toprovide a sufficient image suppression.

• Frequency Tripler. It multiplies the 16 GHz LO frequencyby a factor of three up to 48 GHz, thus generating the LOsignal needed by the V-band active mixer for the final up-conversion.

• V-Band Active Mixer. This circuit performs the final fre-quency conversion that finally brings the data signal in the62 − 63 GHz transmission band using the 48 GHz LO. Itsconversion loss has to be as low as possible.

• 16 and 48 GHz Amplifiers. They are necessary to bring thedata signal and the LO signal to the needed power levels.

• Power Splitter. All the necessary passive components weredesigned and fabricated. They are, in modern microwavetechniques, straightforward to implement and will not bepresented.

• Power Amplifier (PA). The notation PA indicates the am-plifying chain needed to rise the power of the V-band mixeroutput signal up to 10 dBm necessary for transmission.These circuits were not developed during the project.

• Antenna. Antennae were only briefly investigated duringthe project, therefore no result will be presented.

Conclusions 25

2.3 Conclusions

In this chapter, a feasible system architecture for a millimeter-wave fiber-optic communication system was presented and re-alistic specification for the system hub were derived. Remotegeneration of the mm-wave local oscillator through frequencymultiplication in the hub is proposed as the best way to gener-ate a stable high-frequency LO signal and contemporaneouslyovercome the problems due to chromatic dispersion. A two-stepsignal up-conversion in the hub is suggested instead of a directup-conversion in order to relax the mixer specifications in termsof LO- and image-rejection and allow the frequency conversioncircuits to be fabricated on our InP HEMT MMIC process. Inorder to properly set the hub specification the fiber-optic andthe radio links were investigated. For the radio link the limita-tions on the transmitted power due to the current regulationson electromagnetic emission were taken into account. The anal-ysis of the fiber optic link leads to extremely loose specificationson the hub optical front-end in terms of noise figure. Non-lineareffects were, anyway, not considered. The radio link investiga-tion indicates that a transmitted power between 5 and 15 dBmensures a reliable transmission in a 10 m radius and does notviolate law limitations.

3

Process Technology

This chapter describes the InP HEMT process technology em-ployed at the Microwave Electronics Laboratory of the SwissFederal Institute of Technology in Zurich. The first section in-troduces the physics of the high electron mobility transistorand gives an overview of the different material systems usedin HEMT technology. In the second section, the MMICs fabri-cation process is illustrated, while the last part of the chapterpresents the characteristics of the active (HEMTs) and pas-sive (transmission lines, resistors, capacitors, inductors) com-ponents available on the process. Details about passive com-ponents modeling for circuit simulations are also given in thelast section, while InP HEMTs modeling will be specificallyaddressed in the next chapter.

3.1 The High Electron Mobility Transistor

3.1.1 From the two-dimensional electron gas to InP

HEMTs

Since 1978, when the two-dimensional electron gas was discov-ered in GaAs at the n-AlGaAs/GaAs heterointerface [30], andthanks to the development of molecular beam epitaxy (MBE)growing techniques for III-V compound semiconductors, manydifferent heterostructure field-effect transistors with increasingperformance have been demonstrated. The first heterostruc-ture FETs were fabricated by research groups in the UnitedStates [31,32], in Japan [33], and in Europe [34]. These deviceswere given various names: high electron mobility transistor

28 Process Technology

(HEMT), selectively doped heterojunction transistor (SDHT),modulation-doped FET (MODFET), two-dimensional electrongas FET (TEGFET). Ketterson [35] was the first to introducein 1985 a pseudomorphic InGaAs quantum well between the n-AlGaAs/GaAs layers. InGaAs has higher carrier mobility andsuperior saturation velocity than GaAs; this resulted in a de-vice with improved power-gain and noise performance at highfrequencies, better 1/f noise and less generation-recombinationnoise [36]. A couple of years later researchers from different lab-oratories [37–39], taking advantage of the excellent transportcharacteristics of the InGaAs layer, started to use the lattice-matched InGaAs quantum well in the AlInAs/InGaAs struc-ture on an InP substrate. These efforts resulted in one of themost promising high-speed technologies based on III-V com-pound semiconductors.

3.1.2 Physical principles of heterostructures

The basic idea behind HEMTs is creating a 2-DEG in the de-vice channel. In a 2-DEG, free channel electrons are physicallyseparated from the ionized donors, reducing ionized impurityscattering and consequently enhancing electron mobility. A 2-DEG can be obtained when two semiconductors with differentbandgap energies EG,1 and EG,2 and unequal electron affinitiesχ1 and χ2 are interfaced. At thermal equilibrium, the Fermilevels of the two materials align. If the difference between thebandgap energies is greater than the difference between theelectron affinities (EG,1 − EG,2 > χ1 − χ2), a discontinuity inthe conduction band appears at the heterointerface and theconduction band of the small-bandgap semiconductor bendsbelow the Fermi level creating a triangular potential well. Elec-trons are transferred from the material with the highest con-duction band energy into the potential well and, as illustratedby Fig. 3.1, a 2-DEG is created. The electron density in the 2-DEG can be increased by selectively doping the material withthe higher EC , from which the free electrons are donated; thisis called doping modulation. In this way a greater separation

The High Electron Mobility Transistor 29

between electrons and donors is achieved, allowing the 2-DEGto suffer less from Coulomb scattering and to enjoy an highermobility. In order to achieve an even higher electron mobility,the distance between the donors and the 2-DEG can be furtherincreased by means of an undoped spacer layer. This makes themobility of free channel electrons greater than in bulk materi-als. The great transport properties of 2-DEG can be exploitedto fabricate devices excelling for low-noise and high-frequencyperformances.

3.1.3 HEMT Material Systems

The availability of good substrate materials and the quality ofthe epitaxial layers that form the active region of a device areindispensable ingredients for the successful fabrication of het-erostructures and, thus, of HEMTs. The increased substratequality (better orientation mismatch, less defect density) andthe advances in Molecular Beam Epitaxy (MBE) techniques(increased uniformity, better material control) are responsiblefor the enormous progresses of the last decade. High electronmobility transistors can be fabricated using various materi-als. The first HEMTs used a AlGaAs/GaAs heterojunctionon a GaAs substrate, which is easy to fabricate because of

Fig. 3.1 Heterojunction band diagram. ∆EC = χ1 − χ2


Fig. 3.2 Bandgap energy versus lattice constant for various compound semi-

conductors.

the perfect lattice match between AlGaAs and GaAs. Nev-ertheless, the GaAs channel has a limited electron mobilityand of a poor electron confinement. An valid alternative is theInxAl1−xAs/InxGa1−xAs heterojunction. Both electron mobil-ity and electron confinement in the potential well increase withan increasing Indium content in the channel. However, on aGaAs substrate, the strong lattice mismatch between InGaAsand GaAs entangles the growth of good quality layers. The lim-itations of GaAs substrates were circumvented with the intro-duction of metamorphic structures [40,41] in the late nineties.The idea is to distribute the strains deriving from the latticemismatch in a thick buffer layer with an increasing lattice con-stant grown on the GaAs substrate. High-quality heterostruc-tures can be subsequently built on top of the buffer layer. Thistechnique is, unfortunately, still very difficult to domesticateusing MBE. Lattice match and high Indium content in thechannel can be simultaneously obtained using InP as substratematerial. As shown in Fig. 3.2, the Indium content in both In-AlAs and InGaAs can be chosen so that the lattice constants

The High Electron Mobility Transistor 31

Fig. 3.3 HEMT Structure

of the two layers is matched to that of InP (In0.52Al0.48As,In0.53Ga0.47As), thus allowing the growth of lattice matcheddevices on an InP substrate. In this way, room-temperaturemobilities up to 12000 cm2/Vs and electron densities up to3.5 × 1012 cm2 can be obtained.

3.1.4 Device Structure

In Fig. 3.3 the HEMT layer structure used in this work is de-picted. The active layers are grown using molecular beam epi-taxy. The layer geometry and composition were optimized forboth high frequency and low noise applications [42, 43]. Thesubstrate is Fe-doped semi-insulating InP with a thickness of600 µm, on top of which an InAlAs buffer layer with a thick-ness of 300 nm is grown. Its function is to compensate for thesurface roughness of the substrate and to allow the growth ofhigh-quality active layers. The heterojunction is formed by a50 nm thick In0.53Ga0.47As channel, followed by a 10 nm thickIn0.52Al0.48As spacer and by a 15 nm thick Schottky layer ofthe same material. Between the spacer and the Schottky layera Silicon δ-doping plane (2 × 1012 cm−2) supplies the channelwith additional electrons. The top layer is a n+ Silicon-dopedIn0.53Ga0.47As cap layer which protects the Aluminum-rich lay-ers from oxidation and allows the formation of ohmic contacts


with the device channel.The next section explains how, starting from this epi-structure,InP HEMT integrated circuits are fabricated.

3.2 Process Description

A good knowledge of the technology available to the circuitengineer is indispensable to fully understand the design of agiven circuit. This section contains a brief description of theInP HEMT MMIC process developed in the clean room fa-cilities of the Laboratory for Electromagnetic Fields and Mi-crowave Electronics. The complete fabrication process can besummarized in the following steps, as schematically illustratedby Fig. 3.4.

1. The device active layers are grown using molecular beamepitaxy on a semi-insulating Fe-doped InP substrate. Theepitaxial wafers used in this work were bought from com-mercial suppliers [44].

2. The wafer is diced into several 8.5 × 8.0 mm2 chips. Eachchip is processed separately.

3. Ohmic contacts are defined. Ge (17 nm), Au (48 nm), Ni(10 nm), and Au (200 nm) are first evaporated and, after liftoff, annealed for two minutes at 300 C in N2 atmosphere.The ohmic contacts are patterned using an image reversalresist.

4. Isolation between the devices is obtained by mesa etching.The active layers are etched down to the buffer layer usinga H3PO4:H2O2:H2O 1:1:5 solution. The successful deviceisolation is verified with I-V on-chip measurements.

5. The gate-recess of the device is etched in order to form aSchottky contact between the gate and the channel. First,the T-shaped gate is patterned with electron beam lithog-raphy using a three-layers resist stack. After resist devel-opment, the InGaAs cap layer is etched using an highlyselective succinic acid solution. The same solution etchesfrom the side the InGaAs channel which remains exposed

Process Description 33

Fig. 3.4 Schematic representation of the used InP HEMT MMIC fabrication

process.


after device isolation. This prevents contacts between thegate metallization and the channel, thus reducing the gateleakage current [45].

6. The gate metal (20 nm Ti, 250 nm Au) is formed. Titaniumis used to prevent Gold diffusion as well as to improve thegate mechanical stability. The T-shaped gate metal allowsthe fabrication of very short gates with low parasitic gateresistance.

7. Thin film resistors are first patterned using an image-reversalresist and then fabricated by evaporation of 23 nm Ti. Thethickness of the evaporated layer is optimized for 50 Ω/resistors.

8. The first metal interconnect is defined using an image-reversal resist; the evaporated metallization consists of 20nm Ti and 250 nm Au.

9. A SiNx dielectric layer for simultaneous device passivationand MIM capacitors fabrication is deposited. This is doneat Avalon Photonics [46] using plasma-enhanced chemicalvapor deposition and reactive ion etching.

10. The second metal interconnect is defined using an image-reversal resist; the evaporated metallization consists of 20nm Ti and 250 nm Au. It is noticeable that this second met-allization is everywhere coplanar with the first one, exceptin those areas where SiNx is deposited.

11. The air bridge feet are defined using a 4 µm thick positiveresist and 20 nm Ni evaporation.

12. The air bridges are patterned with a 3 µm thick positiveresist and then electroplated with Gold. The thickness ofthe electroplated Gold in 1 µm. Electroplating is also usedto increase the metal thickness of passive components suchas inductors and coplanar wave-guides.

The fabrication process of a complete MMIC is performed inapproximately two weeks. Figure 3.5 shows a detail of a fab-ricated MMICs, where different circuit components such as

Components 35

Fig. 3.5 Detail of a fabricated MMICs. Various circuit components like

HEMTs, capacitors, transmission lines etc. can be identified.

HEMTS, capacitors, transmission lines etc. can be recognized.The characteristics both active and passive components avail-able on this process will be described in the next section.

3.3 Components

This part of the chapter describes the different circuit compo-nents that can be implemented with the process illustrated inthe previous section. Circuit design-oriented modeling of pas-sive components is also presented and comparisons betweenmeasurement and simulations are shown. Large-signal modelingof InP HEMTs will be extensively treated in the next chapter.

3.3.1 InP HEMTs

The standard HEMT device used in this work can be seen inFig. 3.5. It is a two finger transistor with a gate length of 0.2µm and a total gate width of 2×75 µm. The HEMT has a “but-terfly” layout, which presents two important advantages. First,the low parasitic effects simplify high-frequency measurements,


0 0.5 1 1.5 2 2.50

50

100

150

200

250

300

350

400

Vds

(V)

I d (m

A/m

m)

Fig. 3.6 I-V characteristics of the typical HEMT device. Vgs ranges from −1 V

to 0.5 V with 50 mV steps.

and, consequently, full characterization at millimeter-wave fre-quencies. Additionally, the symmetric layout can be easily in-terfaced with transmission lines in a coplanar environment, es-pecially in those circuits where a common-source configurationis needed. Typical I-V curves for these devices are depictedin Fig. 3.6. The maximum saturated drain current lies around350 mA/mm, the maximum transconductance is 580 mS/mm,and the typical threshold voltage VT is −0.5 V. The I-V char-acteristics indicate a perfect pinch-off and a good saturation,with a low DC output conductance. This is an indication of thegood electrical qualities of the process.Figures 3.7 and 3.8 show the extraction of the high-frequency

figures of merits, i.e. the transit frequency fT and the maxi-mum frequency of oscillation fmax. The transit frequency isthe frequency at which the current gain h21 is unity. The maxi-mum frequency of oscillation is defined as the frequency wherethe maximum available gain (MAG) or the Manson’s unilat-eral power gain (U) become unity. In this work we chose toextract fmax from the Manson’s unilateral power gain, but it

Components 37

can be easily demonstrated that both the MAG and U becomeunity at the same frequency [47]. Both fT and fmax are ex-trapolated using a tangent with a −20 dB/decade slope. AnfT of 135 GHz and an fmax of 200 GHz are obtained. The ex-tracted values of the high-frequency figures of merit attest thedevice suitability for V-band application. Another importantfigure of merit is the minimum noise figure (NFmin). Figure3.9 shows NFmin of a typical device as a function of frequency.The device is biased in the saturation region and the bias draincurrent is 10 mA. The low-noise characteristics of these devicesare excellent and comparable to the best reported literatureresults. Cryogenically-cooled low-noise amplifiers with recordNFmin and noise temperature where demonstrated using thisprocess [48].

3.3.2 Transmission Lines

Transmission lines are vital components of any analog millimeter-wave circuit. While at microwave frequencies the possibility of

108

109

1010

1011

0

5

10

15

20

25

30

35

40

45

50

Frequency (Hz)

h 21 (

dB)

Fig. 3.7 Current gain h21 of a typical device and its associated fT , extracted

using a −20 dB/decade tangent.


109

1010

1011

0

5

10

15

20

25

30

35

40

45

50

Frequency (Hz)

Uni

late

ral P

ower

Gai

n (d

B)

Fig. 3.8 Manson’s unilateral power gain U of a typical device and its associated

fmax, extracted using a −20 dB/decade tangent.

integrating lumped elements often represents an efficient al-

2 4 6 8 10 12 14 16 18 200.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

Frequency (GHz)

Fm

in (

dB)

Fig. 3.9 Minimum noise figure as a function of frequency for a typical device.

The transistor is in saturation region and the bias drain current is 10 mA.

Components 39

ternative to transmission lines, no other option is available atmillimeter-wave frequencies. In this work, coplanar waveguides(CPW) are used, which were first proposed by Wen in 1969 [49].A schematic representation of the cross section of a CPW isshown in Fig. 3.10. This structure supports quasi-TEM propa-gation up to a certain frequency, which depends on the ground-to-ground spacing. A rule of thumb is given by:

d < λmax/10, (3.1)

where λmax is the wavelength along the line of the maximumpropagation frequency and d is the ground-to-ground spac-ing [50].Coplanar waveguides were preferred to microstrip lines for var-ious reasons. In CPW technology, no wafer thinning, back-side metallization, and via holes through the InP substrate areneeded. This reduces significantly the needed fabrication steps.Additionally, a via hole to ground introduces a parasitic induc-tance, which can be avoided with the presence of a coplanarground on top of the substrate. Another benefit of coplanarwaveguides, compared to microstrip lines, is the fact that theirdimensions are not univocally defined for a given impedance.This introduces an element of flexibility that can be exploitedby the circuit designer. CPWs can be easily flip-chip bondedon a coplanar substrate, which results in better signal transi-tions between the mounting substrate and the integrated cir-cuit. More details on flip-chip bonding of coplanar circuits willbe given in the last chapter.Employing coplanar waveguides presents also some drawbacks.Losses are generally higher than in microstrip lines. If the groundpotential on both sides of the center conductor is not equal,asymmetric modes might start to propagate in the CPW. Thiscan happen when the travelling signal reaches a discontinuitylike a T-junction, a cross-junction, or a 90 bend. In order toavoid asymmetric modes propagation, the ground planes haveto be short-circuited both in proximity of a discontinuity and


Fig. 3.10 Schematic cross-section of a coplanar waveguide.

along long (> λ/4) transmission lines.The lack of models in commercial circuit simulators repre-

sents, from a designer’s point of view, a significant disadvan-tage. Modelling of CPW discontinuities will be addressed in thenext section.In order to guarantee quasi-TEM propagation up to the V-bandand further, a ground-to-ground spacing of 90 µm was chosenin this work. By changing the ratio between the center conduc-tor width w and the gap g, the line impedance can be varied ina range between 30 and 70 Ω. In order to reduce propagationlosses, the thickness of the center conductor was increased from0.26 to 0.8 µm using electroplating over the first level of metal.Agilent LineCalc was used to determine the line dimensionsfor different impedances, and the CPW model implemented inAgilent ADS was used to model straight transmission lines.

3.3.3 Line Discontinuities

In order to perform reliable simulations of coplanar circuits atmillimeter-wave frequencies, it is mandatory to carefully con-sider the effects of CPW discontinuities such as T-junctions,cross-junctions, 90 bends, short circuits, and open circuits.Figure 3.11 shows some fabricated CPW discontinuities. Asalready mentioned, the simulation of coplanar circuits is com-plicated by the absence, in commercial circuit simulators, ofreliable models for CPW discontinuities. A possible way to cir-cumvent this problem is the approach presented in this sec-tion [42]. With the help of a “3d-planar” [51] electromagneticsimulator the S-parameters of the discontinuities are calculatedover a broad range of frequencies. Since the simulated struc-

Components 41

tures have small sizes, the computational effort is limited. TheS-parameters of the modeled structures are then imported asdata files into the circuit simulator. This method has the ad-vantage of being simple and flexible, but requires some verifi-cations. In order to prove the correctness of the simulations,test structures were fabricated and then measured. Figure 3.12shows the measured and simulated S-parameters of a 1.3 mmfolded CPW with four consecutive 90 bends and ending in anopen circuit. Both the simulated magnitude and phase of theinput return loss are in excellent agreement with the measureddata. Another example is illustrated in Fig. 3.13, which containsthe measured and simulated S-parameters of a 48 GHz ampli-fier containing T-junctions, cross-junctions, open circuit stubs,and short circuit stubs in its matching networks. In the sim-ulation, the measured S-parameters of the active device wereused. The agreement is excellent, and the frequency differencebetween the measured and simulated optimum return losses isonly 2% (49 GHz instead of 48 GHz).

Fig. 3.11 CPW discontinuities. (a) 90 bend, (b) T-junction, (c) cross-junction,

(d) air bridge along a line.


0 10 20 30 40 50 60 70 80−2.5

−2

−1.5

−1

−0.5

0

Frequency (GHz)

S11

(dB

)

0 10 20 30 40 50 60 70 80−200

−100

0

100

200

Frequency (GHz)

Pha

se S

11 (

deg)

Fig. 3.12 Measured (continuous line) and simulated (dashed line) S-parameters

of a 1.3 mm long folded CPW presenting four consecutive 90 bends and ending

in an open circuit.

30 35 40 45 50 55 60−30

−25

−20

−15

−10

−5

0

5

10

Frequency (GHz)

S−

para

met

ers

(dB

)

Fig. 3.13 Measured (symbols) and simulated (continuous line) S-parameters

of a 48 GHz amplifier presenting various T-junctions, cross-junctions, and open

circuit stubs.

Components 43

0 10 20 30 40 50 60 70 80−200

−100

0

100

200

Frequency (GHz)

Pha

se S

21 (

deg)

0 10 20 30 40 50 60 70 80−40

−30

−20

−10

0

Frequency (GHz)

S11

, S21

(dB

)

Fig. 3.14 Measured (continuous line) and simulated (dashed line) S-parameters

of a 1 mm long CPW crossed by 11 air bridges with regular spacing of 100 µm

between each other.

A different approach was used for air bridges. With the pro-cess described in section 3.2, only air bridges with a maximumlength of 50 µm can be reliably fabricated. Since the ground toground spacing of the typical CPW is 90 µm, i.e. to wide for anair bridge, the first level of metallization is used to short circuitthe ground planes of the coplanar waveguides and the air bridgeis fabricated on the center conductor, as it can be seen in Fig.3.11. Air bridge effects at T-junctions, cross-junctions, and 90

bends are included in the electromagnetic simulations, but airbridges along a straight line have to be taken separately into ac-count. A capacitor to ground was found to be a suitable model;3 fF is the typical capacitance value for an air bridge on a 50Ω transmission line with 90 µm ground to ground spacing. Fig-ure 3.14 shows the measured and simulated S-parameters of a1 mm long coplanar waveguide crossed by 11 regularly spacedair bridges. The good agreement between measurements andsimulations confirms the validity of the modeling approach.


3.3.4 Thin Film Resistors

Thin film resistors are fabricated by evaporation of a Titaniumlayer with a thickness of 23 nm. The resistive layer has a resis-tance of 50 Ω/ with a typical variation of 3% over the samechip and less than 10% tolerance over different chips. It is worthmentioning that, in order to protect the Titanium layer dur-ing the RIE step performed after dielectric deposition, the thinfilm resistors are protected by a SiNx passivation layer. Whendrawing the layout, the designer must ensure that the contactsbetween resistors and metal are done using the first metalliza-tion layer. If the resistors are covered with passivation, in fact,it is not possible to contact them with the second metallizationlayer.

3.3.5 MIM Capacitors

On this process, metal-insulator-metal (MIM) capacitors arefabricated using the first metallization level as bottom plate,a 120 nm thick SiNx layer as dielectric and the second metal-lization layer as top plate. In this way, capacitors with a ca-pacitance of 0.5 fF/µm2 can be fabricated. For reasons rangingfrom mask tolerance to alignment problems and lithographyresolution issues, the fabrication of MIM capacitors with capac-ity values lower than 20 fF is not recommended. Inter-digitalcoplanar structures are better suited for this purpose. The pro-cess tolerance, which depends on the variations of the dielectricthickness and dielectric constant is less than 10% for differentchips, and less than 3% on the same chip [52].

3.3.6 Spiral Inductors

At Ku-band frequencies and lower, the large physical dimen-sions of transmission lines make their employment in the fabri-cation of passive structures not practical anymore. If a small cir-cuit size is needed, the use of lumped inductors becomes neces-sary. During this work, spiral inductors with different sizes andgeometries were designed, fabricated, and characterized. Tables3.1 and 3.2 provide an overview of the fabricated inductors with

Components 45

Table 3.1 Square spiral inductors fabricated with the InP HEMT process.

Conductor width and conductor spacing is 10 µm. Metallization thickness is

0.8 µm.

Type Inductance (nH) DC Resistance (Ω) Qmax fsr (GHz)

L250 0.35 3.2 13.3 > 40L600 0.75 4.2 9.8 36L1000 1.04 4.86 9.8 32.4L1300 1.2 4.8 8.76 25.6L2000 1.9 6 10.05 21.5L3500 3.9 10.6 7.09 10.7

Table 3.2 Circular spiral inductors fabricated with the InP HEMT process.

Conductor width and conductor spacing is 10 µm. Metallization thickness is

0.8 µm.

Type Inductance (nH) DC Resistance (Ω) Qmax fsr (GHz)

L530 0.65 4.2 10.6 40L750 0.9 4.6 10 31L1010 1.3 5.4 9.5 24.7L1310 1.6 6.2 8.8 22.3L1550 2.1 7.2 8.6 18.2L1860 2.4 7.7 8.6 16.6L2020 2.6 8.3 8.5 15.6L2630 3.6 7.9 9.9 13L3200 4 10.7 7.5 11.7L3360 4.3 11.3 7.6 11.2

their most important parameters, i.e. inductance value, DC re-sistance, quality factor, and self-resonance frequency fsr. Twoinductor families were developed, one with a square spiral ge-ometry and the other with a round spiral geometry. In bothcases, the conductor width is 10 µm, the conductor spacing is10 µm, and the metallization thickness is 0.8 µm. Their induc-tance ranges from 0.35 to 4.3 nH, and the quality factor Q variesbetween 7 and 13. The low quality factor is due to the limitedmetallization thickness, which results in a high series resistanceand, thus, in high losses. Nevertheless, the developed inductors


were successfully employed in circuit fabrication. No inductormodel was developed; the measured S-parameter were directlyimported in the circuit simulator as data files.

4

InP HEMT Large SignalModel

The development of the millimeter-wave fiber-optic communi-cation system proposed in the second chapter embraces thedesign and the fabrication of a variety of linear and non-linearcircuits with operating frequencies ranging from a few GHz tothe V-band. It is, therefore, important to carefully model thelinear and the non-linear characteristics of the employed devicesboth in the lower GHz range and at millimeter-wave frequen-cies.For InP HEMTs, one of the most important effects at low fre-quencies is the so called kink effect related to impact ioniza-tion [53]. Impact ionization results in high gate leakage current,high output conductance, and low breakdown voltage. Due tothe smaller bandgap of the InGaAs channel of InP HEMTs, thisphenomenon occurs at lower drain-source voltages compared toGaAs-based devices. Its effects, in fact, can be observed at op-erating conditions typical for non-linear circuits and should,therefore, be taken into account by large-signal models. In theframework of the MFOCS project, an efficient method to ex-tract a large-signal look-up table based model for InP HEMTswas developed that takes impact ionization into account [54].The large-signal model has been verified with linear measure-ments up to the W-band and with non-linear measurements upto the V-band.This chapter is organized as follows; the first section gives

48 InP HEMT Large Signal Model

an overview of the different modeling approaches and brieflypresents the state of the art of large-signal FET modeling. Inthe second section, the used small-signal and large-signal circuittopologies are introduced and discussed. The third section de-scribes the extraction of the small-signal model parameters andtheir integration in the large-signal model. Then, small-signaland large-signal model validation is addressed and comparisonsbetween measured and simulated large-signal performances offabricated MMICs are presented. Finally, a discussion on thelimitations of look-up table based models concludes the chap-ter.

4.1 Large-Signal Modeling of High-Frequency FETs

There are three possible approaches to high-frequency modelingof FET devices. Each of them has different advantages anddisadvantages, and the choice of a specific approach stronglydepends on the application the model is needed for.

1. Physical models describe the device behaviour in termsof carrier transport properties, device geometry, and ma-terial characteristics. This kind of model usually providesa deep understanding of the device, but is inherently com-plex and often requires sophisticated numerical methods toconverge to a solution. Physical models can be efficientlyused for device optimization, but, at the moment, they arenot suited for circuit simulation.

2. Black box models are the conceptual opposite of phys-ical models, since they try to represent a device with aninput-output transfer function, without giving any insighton their nature. The mathematical description of the trans-fer function relies on model parameter which are extractedby fitting the measurement results. Parameter fitting oftenresults in an arduous optimization problem, which makesthe extraction of black box models rather complicated.

3. Small-signal equivalent circuit based models are themost used for circuit simulations. In this kind of model, the

Large-Signal Modeling of High-Frequency FETs 49

device behaviour is described by means of lumped circuitelements, which are extracted from DC and S-parametermeasurements. If the small-signal equivalent topology isproperly chosen, this approach is physically meaningful, ac-curate over a wide range of frequencies, and easy to imple-ment in a circuit simulator.

Since the large-signal model described in this work was neededonly for circuit simulations, it was mandatory to choose a small-signal equivalent circuit based model. Among the many possi-ble implementations of this kind of model, two main familiescan be recognized, which differentiate themselves depending onhow the constitutive relations of the large-signal model are ex-pressed.In empirical models, the large-signal behavior of the deviceis described by means of analytical functions of the terminalvoltages. These functions must fulfill various conditions and areusually fitted on the small-signal parameters extracted from themeasured S-parameters. Empirical models need a small amountof measurements for their extraction 1, but require an effort inchoosing the most appropriate analytical functions and in fit-ting them with the small signal parameters. Empirical modelshave the advantage of allowing reliable device simulations alsooutside the measurement space, but show a limited adaptabil-ity to changes in the device due, for example, to technologyvariations.In a look-up table based model the constitutive relations aretabulated as a function of the terminal voltages. During non-linear simulation the table files are accessed and their values in-terpolated. The large-signal constitutive relations are generatedby numeric integration of the small-signal circuit elements ex-tracted from S-parameter measurements. In order to guaranteea good model accuracy, the interpolation errors must be keptas low as possible. This means that the number of S-parameter

1Normally, for the extraction of an empirical based model, it is sufficient to performS-parameter measurements at 30 different bias points.


measurements needed for the extraction of such a model is gen-erally an order of magnitude higher than for empirical models.An advantage of look-up table based models is their ability ofbeing easily adaptable to device variations, given the invarianceof the small signal topology. The choice of a table based modelis, therefore, the most appropriate for a technology aimed todevice optimization and consequently open to process modifi-cations.Many works on high-frequency FET device modeling have al-ready been reported. Root presented one of the the first look-up table based GaAs FET large-signal models in 1991 [55], butlimited its validation to microwave frequencies. In 1999, as arefinement of Root’s work, Wei presented a table-based modelwith better S-parameters prediction capability [56]; this modelwas also verified only up to microwave frequencies. A MODFETsmall-signal equivalent circuit validated up to 120 GHz was pre-sented by Tasker in 1995 [57], while Wood and Root publisheda bias-dependent scalable model validated up to the same fre-quency in 2000 [58]. Small-signal and noise behavior of impactionization were modeled by Reuter in 1997, but were not given alarge-signal representation. A large-signal model validated upto the V-band with non-linear measurements and up to 118GHz with linear measurements was presented by Fernandez-Barciela in 2000, but this model addresses GaAs FETs and,therefore, does not take impact ionization into account [59].

4.2 Small-Signal and Large-Signal Equivalent Circuit

The devices used for model extraction are two-fingers InP HEMTswith 0.2 µm gate length, 75 µm gate-finger width, 135 GHz fT ,and 200 GHz fmax. The fabrication process and the most im-portant characteristics of these devices were presented in thethird chapter. The modeling procedure is an extension of themethod presented in [60], developed at the K.U.Leuven, andsuccessfully used for GaAs PHEMTs. In [60], where only ex-perimental results up to microwave frequencies are shown, a

Small-Signal and Large-Signal Equivalent Circuit 51

non-quasi-static small-signal model and its corresponding large-signal model are presented, where in the small-signal equivalentcircuit the feedback elements between the gate and the drainterminals are eliminated and replaced with transelements. Thishas the advantage of allowing the small-signal topology to beconsistent with the large-signal equivalent circuit, which helpsto respect charge conservation. For mm-wave effects and impactionization to be included into the model, some modificationshave to be made both to the small-signal and to the large-signal equivalent schemes. The small-signal equivalent circuitis depicted in Fig. 4.1. At mm-wave frequencies the electricallength of gate and drain metallizations becomes relevant andtransmission-line effects need to be included. This is achievedby splitting the extrinsic gate and drain capacitances Cpg andCpd in two [57], [58]. Because of the device layout, distributedeffects on the source metallization can be ignored. Since Cpg

and Cpd are bias-independent extrinsic elements, this splittingdoes not influence the large-signal equivalent scheme.In order to model impact ionization effects, an extra networkis added at the drain side of the small-signal equivalent cir-cuit [61]. This network consists of a resistance Rim in series withthe parallel connection of a capacitance Cim and a transconduc-tance gim controlled by the intrinsic drain-gate voltage Vdgi.The used large-signal intrinsic representation of the transis-

tor is shown in Fig. 4.2. It consists of the parallel connectionof a charge source and a current source at the drain side andof a current source in parallel with the series connection of acharge source and a resistor at the gate side. This large-signalequivalent circuit is consistent with the intrinsic small-signalequivalent scheme already introduced. To evaluate the effectsof impact ionization in non-linear circuits, it is necessary to finda large-signal representation of the small-signal network intro-duced to model it. From experimental results, it was found thatimpact ionization can be best modeled by means of a first or-der dispersive network. This is somehow similar to the modeling

52

InP

HEM

TLarg

eSig

nalM

odel

Ri

Rgdf

Rds

Rim

(G +j C )m mw Vgsi

Ls

Cgi

Rs

source

Rgsf

RgLg

gate

intrinsic

C /2pg Cds

Cim

gimVdgi

j Cw dmVdsi

V’gsi

+ Rd Ld

drain

Vdsi

+

-

Vdgi +-

C /2pg

+

-

Vgsi

C /2pd C /2pd

Fig

.4.1

Sm

all-sig

naleq

uiva

lent

circuit.

Sm

all-S

ignaland

Larg

e-Sig

nalEquiv

alen

tCircu

it53

source

drain

Vdsi

gate

I (V’ ,V )gs gsi dsi

V’gsi

C

Cgi

dm

(V ,V ) dV /dt -

(V ,V ) dV /dtgsi dsi gsi

gsi dsi dsi

Ri(V ,V )gsi dsi

Vgsi C

Cm

ds

(V ,V ) dV /dt +

(V ,V ) dV /dtgsi dsi gsi

gsi dsi dsi

IdsDC LF

LF LF dsLF io

io io dsHF

(V ,V )/(1+j )+

j /(1+j )(I /(1+j )+

j /(1+j )I )

’gsi dsi wt

wt wt wt

wt wt

(V ,V )

(V ,V )

gsi dsi

gsi dsi

Fig

.4.2

Larg

e-signaleq

uiva

lent

circuit.


approach of the dispersive behavior of Gm and Gds in the kHz-MHz range [62]. The bias-dependent values of Gm and Gds areextracted twice, once at frequencies lower than those of inter-est for impact ionization (250 MHz-0.5 GHz), the second timeat frequencies where impact ionization does not influence theS-parameters anymore (7-110 GHz). It must be noticed thatat 250 MHz, which is the lowest measurement frequency, thelow-frequency output dispersion is no longer effective; the onlydispersion effect is the one originated by impact ionization.The low frequency drain-source current IdsLF and the high fre-quency drain current IdsHF are obtained from the integrationof Gm and Gds towards their intrinsic terminal voltages. An-other point worth noticing is that the charges are not obtainedfrom the numerical integration of the small-signal capacitances,but are expressed as analytical functions of the bias-dependentcapacitances themselves, as already successfully demonstratedin [63, 64]. At the gate side this is mathematically equivalentsince terminal charge conservation is guaranteed, neverthelessone should be aware of possible peculiarities. The average ca-pacitive current, in fact, might not be zero, leading to non con-servation of energy [65]. At the drain side terminal charge con-servation is not necessarily guaranteed since two capacitanceswith two different physical meanings are used; Cds is a space-charge capacitance, Cm is related to the intrinsic time delay τ .It is important to underline that terminal charge conservationis different from physical charge conservation, which is alwaysvalid and expressed in circuits theory by Kirchhoff’s CurrentLaw. Terminal charge conservation is related to the bias de-pendence of different capacitance functions attached to a sin-gle node, and implies that all the capacitances attached to agiven terminal can be uniquely described by the partial deriva-tives of a single charge function. The choice of a large-signalmodel formulation which respects terminal charge conservationis simply a modeling choice, which is needed to avoid possibleconsequences, such as a DC current spectral component under

Model Extraction 55

large-signal conditions [66]. Kirchhoff’s laws are always valid ina large-signal model, charge conservation not necessarily.The large-signal equivalent circuit is complete after adding an-other first order transfer function which regulates the transitionbetween the measured DC current IdsDC and the low frequencycurrent IdsLF [62], thus modeling the low frequency output dis-persion. In the next section it will be shown how the small-signal and the large-signal model parameters are extracted fromthe measurements.

4.3 Model Extraction

The first step of the modeling procedure is the extraction ofthe bias-independent extrinsic parameters. This is done fromcold (Vds = 0) measurements using the method described in[67]. The main difference between this method and the classicDambrine approach [68] lies in the extrinsic resistances extrac-tion. The maximum positive value of the gate-source voltageVgs is set to be just below the turn-on of the Schottky gatediode and the asymptotic behavior of Re(Z22) for Rch −→ 0is calculated using Pla’s approximation [69]. In this way, theextrinsic resistances can be calculated while the gate-currentdensity is kept low enough to avoid a possible device degrada-tion.Then, the device S-parameters are measured over a grid of morethan 900 different bias points. Here it is worth pointing outthat in a look-up table based large-signal model the capabilityof predicting high order harmonics is limited by the precisionof the interpolation between those points where the model isactually defined. Since the most important non-linearity is dueto the drain-source current, it was decided to use variable stepsfor Vgs and Vds and have a tighter grid where the drain-sourcecurrent changes more drastically, i.e. close to the threshold volt-age for Vgs and in the linear region for Vds.In order to obtain a good accuracy up to mm-wave frequen-cies, it is not enough to measure the S-parameters at only one


frequency point [55]. In this case it would be also impossibleto have any information about the impact ionization effects,which are strongly frequency dependent. On the other hand,in order to achieve a sufficient resolution in the low frequencyrange with a linear sweep, at least 201 frequency points wouldbe necessary. This would generate a huge amount of data, notto mention the long time required by the measurements. Whatis needed is a good resolution at low frequencies, to extract theimpact ionization related parameters, and only few points inthe mm-wave frequency range, which are sufficient to extractthe other parameters. Therefore, we chose to use a logarithmicfrequency sweep that reaches the goal using only 59 points.This is important because it improves the measurement speedand drastically reduces the computation time required for thesmall-signal parameters extraction.After measurements, the bias-independent extrinsic parame-ters are de-embedded from S-parameters. At this point, thecorrect determination of the extrinsic parameters is crucial fora successful model extraction. Small errors in the extrinsic pa-rameters extraction, in fact, can cause large inaccuracies of thede-embedded intrinsic parameters, which might result in aninvalid model. The intrinsic small-signal circuit elements arecalculated according to the following expressions.

Gds = <(Y22) (4.1)

Gm = <(Y21 − Y12) (4.2)

Ri = <(1/(Y11 − Y11DC)) (4.3)

Cgi = −1/(2πf =(1/Y11)) (4.4)

Cdm = Cgi

∣

∣

∣

∣

Y12

Y11

∣

∣

∣

∣

(4.5)

Cds = −1/(2πf =(1/Y22)) (4.6)

Cm = =(Y21 − Y12)/(2πf) (4.7)

Model Extraction 57

τio =1

f |min(Y22)

(4.8)

In equation 4.3 Y11DC is subtracted from Y11 because the Igs

state function is directly measured at DC. This is allowed by thenon-dispersive behavior of Igs. The values of Gds and Gm lead-ing to IdsLF are extracted at the lowest measurable frequency(250 MHz). Gds used in IdsHF is extracted at a frequency equalto 1/τio, where τio is the time constant regulating impact ion-ization dispersion. Gm used in IdsHF and all the other intrinsiccomponents can be well determined by taking their mean valueover the 10 GHz - 25 GHz range, where no dispersive effecttakes place.A possible degradation of the device due to the multi-bias S-parameter measurements is checked by comparison of the DCand RF characteristics before and after the measurement pro-cess [70]. This is done to ensure the validity of the extrac-tion procedure. All the measurements and the extraction ofthe small-signal parameters are performed using Agilent IC-CAP. The device measurements last about 3 hours, while thede-embedding procedure and the parameters extraction needapproximately one minute on a SUN workstation. After thesmall-signal parameter extraction, integration is performed us-ing numerical routines implemented in MATLAB.Here, a final problem has to be solved. For the model to behandled in a non-linear circuit simulator as Agilent ADS, theintrinsic terminal voltages Vgsi and Vdsi contained in the look-up tables must lie on an orthogonal grid. At the same time alsothe extrinsic terminal voltages Vgse and Vdse need to be on anorthogonal grid, otherwise one would have to define one by oneall the different bias points used in device measurements. Theintrinsic and the extrinsic terminal voltages are related by thefollowing equations.

Vdsi = Vdse − (Rdc2 + Rd + Rs)Ids(Vgse, Vdse)

− RsIgs(Vdse, Vgse)(4.9)


Vgsi = Vgse − (Rdc1 + Rg + Rs)Igs(Vgse, Vdse)

− RsIds(Vdse, Vgse)(4.10)

In equations (9) and (10) Rdc1 and Rdc2 are the resistancesof the cables used to bias the device and Rg, Rd, Rs are theextrinsic gate, drain and source resistance, respectively. From(9) and (10) it is clear that measuring using an orthogonalextrinsic-voltages grid does not lead to an orthogonal intrinsic-voltages grid. The problem was solved by generating an orthog-onal intrinsic-voltages grid and interpolating the values of thesmall-signal intrinsic parameters expressed as a function of thenon-orthogonal intrinsic-voltages on the orthogonal intrinsic-voltages grid. In this way, the small-signal intrinsic parameterscan be integrated on an orthogonal grid. As it is shown in thenext section, this interpolation does not affect the model pre-cision.After numerical integration, the tabulated values of the large-signal constitutive relations are saved as CITI files. The modelis finally implemented in the Agilent ADS circuit simulator asSymbolically-Defined Device.

4.4 Model Validation

The accuracy and limitations of the model were extensivelytested under both small-signal and large-signal conditions. Themost significant results are reported in the next sections.

4.4.1 Small-signal verification

A comparison between measured and simulated S-parametersis shown in Fig. 4.3, 4.4, 4.5, and 4.6. The device is biased atVgs = −0.2 V, Vds = 1.4 V. The bias point was deliberatelychosen to be in a region where impact ionization is affectingthe device characteristics, as it can be noticed by the reductionin the power gain S21, and by the inductive behavior of theoutput return loss S22, both at low frequencies. In order tobetter visualize the effect of impact ionization S21 is shownboth on a polar plot and on a dB scale. (Figures 4.5 and 4.6).

Model Validation 59

S22

S11

− simulation ° measurement

Freq 0.5 to 110 GHz

Fig. 4.3 Comparison between measured (symbols)and simulated (solid line)

S-parameters at (Vgs, Vds)=(−0.2 V, 1.4 V).

0.15 0

S12


Freq 0.5 to 110 GHz




50

S21


Freq. 0.5 to 110 GHz



0 20 40 60 80 100 120−6

−4

−2

0

2

4

6

8

10

12

14

Frequency (GHz)

S21

(dB

)

S21



Model Validation 61

Figures 4.3, 4.4, 4.5, and 4.6 show an excellent agreementbetween measurements and simulations from 250 MHz to 110GHz, and evidence how well the kink effect is reproduced. Thiscorroborates the used small-signal modeling approach and con-firms that a first-order transfer function governing the transi-tion between a low-frequency drain current IdsLF a the high-frequency drain current IdsHF effectively represents the disper-sive device behavior due to impact ionization. Comparisonsof measured and simulated device S-parameters in differentoperating regions were also performed; good results were ob-tained [71], fully validating the small-signal prediction capabil-ity of the developed model.

4.4.2 Large-signal verification

After the small-signal verification, the model was tested un-der non-linear operating conditions. In Figure 4.7 a compari-son between measured and simulated third-order intermodula-tion product is shown. The measurements were performed ina frequency region (3.5 GHz) and at bias point (Vgs = −0.2V, Vds = 1.4 V) where impact ionization is affecting the de-vice. Measurements and simulations are in excellent agreement,and the output power of the fundamental tone as well as thethird-order intermodulation product are well predicted. Theapproach followed to include impact ionization into the modelis, therefore, validated for large-signal conditions as well.Additional non-linear measurements were performed to testthe model capability in predicting the large-signal behavior ofthe device at millimeter-wave frequencies. Figure 4.8 shows themeasured and simulated first, second, and third output powerharmonics for an input frequency of 16 GHz. The device wasbiased at a bias point typical for non-linear circuits such asmixers or frequency multipliers (Vgs = −0.3 V, Vds = 1 V).As we can see the prediction capability of the model is accu-rate up to the third harmonic (48 GHz). This result clearlyindicates that the interpolation method used to generate thelook-up tables does not have a negative impact on the model


precision. An excellent agreement between measurements andsimulations can be noticed also in Fig. 4.9, where the measuredand the calculated first and second harmonics for a fundamen-tal frequency of 32 GHz are shown. In Figure 4.10 the measuredand simulated output power at a fundamental frequency of 64GHz are shown for three different DC bias points. The first

−24 −22 −20 −18 −16 −14 −12 −10 −8 −6 −4−70

−60

−50

−40

−30

−20

−10

0

10

Pin

(dBm)

Pou

t (dB

m)

f1

2f1−f

2

Fig. 4.7 Comparison between measured (symbols) and simulated (solid line)

third-order intermodulation. f1 = 3.5 GHz, f2 = 3.501 GHz. Pin represents the

power of a single tone. The DC bias point is: (Vgs, Vds)=(−0.2 V, 1.4 V).

bias point (a) corresponds to class A operation (Vgs = −0.2 V,Vds = 1 V); in the second (b) Vgs is close to VT and Vds is inthe saturation region (Vgs = −0.4 V, Vds = 1 V); in the third(c) Vgs is close to VT and Vds is in the linear region (Vgs = −0.4V, Vds = 0.5 V).The results presented in this section cover the entire spectrumof possible applications for the developed large-signal model,both in terms of frequency range and device class of operation.The model capability of predicting the non-linear performanceof the used InP HEMTs has, in fact, been verified from im-pact ionization frequencies (3.5 GHz) to the V-band (64 GHz)

Model Validation 63

−10 −5 0 5−50

−40

−30

−20

−10

0

10

Pin

(dBm)

Pou

t (dB

m)

f0

2f0

3f0

Fig. 4.8 Measurements (symbols) and simulated (solid line) output power ver-

sus input power at the fundamental frequency of 16 GHz and the corresponding

second and third harmonics. The DC bias point is (Vgs, Vds)=(−0.3 V, 1 V)

−16 −14 −12 −10 −8 −6 −4 −2 0 2−50

−40

−30

−20

−10

0

10

Pin

(dBm)

Pou

t (dB

m)

f0

2f0

Fig. 4.9 Measurements (symbols) and simulated (solid line) output power ver-

sus input power at the fundamental frequency of 32 GHz and the corresponding

second harmonic. The DC bias point is (Vgs, Vds)=(−0.3 V, 0.7 V)


−25 −20 −15 −10 −5 0−40

−35

−30

−25

−20

−15

−10

−5

0

Pin

(dBm)

Pou

t (dB

m)

a

b

c

Fig. 4.10 Measured (symbols) and simulated (solid line) output power versus

input power at 64 GHz at three different bias points: (Vgs, Vds)=(−0.2 V, 1

V)(a),(−0.4 V, 1 V)(b),(−0.4 V, 0.5 V)(c).

as well as for class A and class AB operation. In the next sec-tion, an insight will be given on the accuracy of the large-signalmodel in predicting the performances of real-world circuits suchas mixers and amplifiers.

4.4.3 Large-Signal Validation through MMIC Design

As already mentioned, the large-signal model discussed in theprevious sections was used to design a large number of circuits.This section presents, as additional large-signal verification, thecomparisons between measurements and simulations for a 64GHz active mixer, a 16 GHz amplifier, and a 64 GHz amplifier.All the circuits were fabricated using the process described inChapter 3.

64 GHz Mixer

The mixer is an up-converting mixer that uses a 48 GHz localoscillator to convert a signal in the 16 GHz band up to 64 GHz.The LO and the IF are both applied at the same port using

Model Validation 65

Fig. 4.11 Photograph of the 64 GHz up-converting active mixer. Circuit size

is 1.6×3 mm2.

an external diplexer. Figure 4.11 shows a photograph of thefabricated circuit. Measured and simulated conversion gain asa function of frequency for a LO power of −5 dBm, and a IFpower of −15 dBm are shown in Figure 4.12. The measuredvalues show a relatively large spread. Nevertheless, the maxi-

61 61.5 62 62.5 63 63.5 64 64.5 65 65.5−10

−9

−8

−7

−6

−5

−4

Frequency (GHz)

Con

vers

ion

Gai

n (d

B)

Fig. 4.12 Measured (symbols) and simulated (solid line) conversion gain of the

62 GHz active mixer.


Fig. 4.13 Photograph of the 16 GHz amplifier. Circuit size is 1.4×1 mm2.

mum difference between measurements and circuit simulationsis 1 dB at 64 GHz.

−10 −8 −6 −4 −2 0−10

−8

−6

−4

−2

0

2

4

6

8

10

Pin

(dBm)

Pou

t (dB

m)

1dB

Fig. 4.14 Measured (symbols) and simulated (solid line) 1 dB compression

point of the 16 GHz amplifier.

Model Validation 67

Fig. 4.15 Photograph of the 64 GHz amplifier. Circuit size is 1.2×0.54 mm2.

The circuit is biased through external bias-tees.

16 GHz Amplifier

In Fig. 4.13 a photograph of the 16 GHz amplifier is repre-sented. Input and output matching are realized using lumpedelements and unconditional stability is obtained by inductivelyloading the device source. Figure 4.14 reports the measuredand simulated 1 dB compression point. The circuit has a lineargain of 9.2 dB and reaches the 1 dB compression point an in-

−18 −16 −14 −12 −10 −8 −6 −4 −2 0 2−20

−15

−10

−5

0

5

10

Pou

t (dB

m)

Pin

(dBm)

1dB

Fig. 4.16 Measured (symbols) and simulated (solid line) 1 dB compression

point of the 64 GHz amplifier.


put power of −0.5 dBm. The behavior of the circuit could benicely reproduced by the simulations. Higher order harmonicswere not measured.

64 GHz Amplifier

The 64 GHz amplifier depicted in Fig. 4.15 uses single-stubmatching both at the input and at the output of the HEMT.Since no bias network is present, the active device is biasedthrough the measurement probes by means of external bias-tees. Measurements and simulations of its 1 dB compressionpoint are reported in Fig. 4.16. At 64 GHz, the circuit shows alinear gain of 5 dB and gets into compression for 0 dBm inputpower. Also in this case, a good agreement was found betweenmeasurements and simulation results.

4.5 Limitations of Look-up Table Based

Large-Signal Models

Look-up table based large-signal models represent a very at-tractive solution in all those situations where a long term pro-cess stability can not be guaranteed, which is the case in aca-demic and industrial research laboratories. Developing an em-pirical model is, in fact, not affordable if the functions describ-ing the large-signal device behavior need continuous modifica-tions. On the other hand, using a table-based approach a devicewith the same small-signal equivalent topology, but with dif-ferent characteristics (e.g. an InP HEMT with the same activelayers but with a different gate length), can be, with minor ad-justments, modeled using the same procedure described in theprevious sections.The look-up table based approach presents an intrinsic limita-tion that must be carefully considered in order to get correctsimulation results. The model is valid only for those valuesof Vgs and Vds that lie inside the measurement space, i.e. onlywhere the look-up tables are defined. While the constitutive re-lations are interpolated inside the measurement range, outsideof the look-up tables extrapolation is used. This often leads to

Conclusions 69

unreliable results. Consequently, during the simulation of cir-cuits operating close to the borders of the measurements space,special attention must be paid to ensure that the values as-sumed by Vgs and Vds stay inside the region where the look-uptables are defined. This represents a disadvantage compared toempirical models, for which analytical functions can be foundthat satisfactorily approximate the device behavior also outsideof the measurement region.The collected experience leads the author to the conclusionthat an empirical model is generally preferable if process sta-bility can be guaranteed; if this is not possible a look-up tablebased model ensures, with the described limitation, an excellentflexibility together with a very good precision.

4.6 Conclusions

This chapter describes an efficient procedure to extract a look-up table-based InP HEMTs large-signal model which takes im-pact ionization effects into account. A significant reduction ofmeasurement and extraction times is obtained by using a log-arithmic frequency sweep. S-parameter measurements demon-strate that the linear performance of the model is excellent upto 110 GHz. The non-linear performance is validated in the im-pact ionization region as well as up to 64 GHz by non-linearmeasurements at different bias points. It is also shown that thedescribed large-signal model can be employed to successfullypredict the performances of amplifiers and mixers up to theV-band.

5

MFOCS Circuits

This chapter presents those integrated circuits, developed inthe frame of the MFOCS project, which are the most interest-ing for a circuit designer, the most critical for the overall systemperformance, and the most scientifically relevant as well. Theyinclude the 0 − 20 GHz optical receiver amplifier, the two up-converting mixers, and the frequency multiplier. For each cir-cuit, both the design approach and the measured performancewill discussed in details.

5.1 0-20 GHz Traveling Wave Amplifier

5.1.1 Introduction

In the MFOCS system, the signal sent from the base station tothe hub through the optical fiber consists of data part (1 − 2GHz), which contains the information, and of a 16 GHz carrier,which is used in the hub to generate the mm-wave local oscil-lator. The light signal is received and converted to an electricalsignal by means of a photo diode. In order to keep the huboverall noise figure as low as possible, the electrical signal gen-erated by the photo diode needs to be immediately amplifiedbefore any signal processing is performed. For this purpose, abroadband 0 − 20 GHz amplifier was designed which amplifiesboth spectral components of the received signal. Only after am-plification the data part will be separated from the carrier bymeans of a frequency selective diplexer. The used design ap-proach and the measurement results are described in the nexttwo sections.

72 MFOCS Circuits

Fig. 5.1 Small-signal representation of the 0-20 GHz traveling-wave amplifier.

5.1.2 Circuit Design

There are two different methods for designing broadband base-band amplifiers: a “lumped elements” configuration or a “dis-tributed” traveling-wave solution. The first approach impliesthe use of feedback amplifiers combined with equalizing stagesoutside or within the feedback loop. In this way, low noise fig-ure, high gain, and reasonably small physical dimensions can beobtained. These configurations have some limitations in termsof input and output return losses; it is usually quite difficult toget a good matching over a broad band without using resistors,which lower the gain and increase the noise figure as well. If lownoise figure, high gain, and small physical dimensions are notan issue, the natural choice is to design a traveling-wave ampli-fier (TWA), which has the advantage of presenting a very goodmatching and a flat frequency response over a wide frequencyband.The typical traveling-wave amplifier consists of an input trans-mission line terminated with a resistor, equally spaced gaincells, and an output transmission line also terminated with aresistor. Both the input and the output resistors are matched tothe impedance of the corresponding transmission lines. The in-put signal is applied at the input line terminal and travels down

0-20 GHz Traveling Wave Amplifier 73

to its terminated end, where it is absorbed by the resistor. Asthe voltage wave propagates along the input line, it excites thedifferent gain cells which amplify the signal and transmit it tothe output transmission line, where two voltage waves are gen-erated in two opposite directions. If the phase velocities of theinput and output line are equal, the voltage waves traveling inthe forward direction (i.e. in the direction of the output ter-minal) in the output line add in phase. The signals travelingin the opposite direction do not necessarily add in phase andtheir remaining components are absorbed by the output linetermination. Phase velocities in the input and output lines aredetermined by the line impedance, by the line length, and bythe input and output admittance of the gain cell. The frequencyresponse of a traveling-wave amplifier is limited by the losses inthe input and output transmission lines, which are determinedboth by the non-zero resistance of the line metallizations andby the resistive parts of the input and output impedance of thegain cells. The small-signal circuit schematic of the designedTWA is depicted in Fig. 5.1. It is a typical traveling-wave am-plifier configuration [72–76], with some modifications in orderto achieve unconditional stability and a flat frequency response.For this design, four InP HEMTs in common-source configura-tion were chosen, being four the number of gain cells whichgives the best compromise between gain and bandwidth [75].In order to achieve unconditional stability, small (Rs = 4 Ω)resistors had to be added at the device gate terminals. Theseresistors lower the amplifier gain and increase the noise figure.A “noiseless” alternative could have been to inductively loadthe HEMT sources, but this would have had serious drawbacksin terms of circuit layout. The termination resistances were notmatched to the line impedances, but were increased by 15%;this results in higher return losses, but allows to obtain a flatterfrequency response in the band of interest. In order to achievesize reduction, the transmission lines were folded; the result-ing additional losses were minimized using round bends. Two

74 MFOCS Circuits

Fig. 5.2 Photograph of the traveling-wave amplifier which uses an external bias

network. The circuit size is 2.2× 2 mm2. The lower part of the figure shows the

layout details in proximity of a common-source amplification cell.

different versions of the same amplifier were fabricated, whichhave the same small-signal schematic but differ for the biasing

0-20 GHz Traveling Wave Amplifier 75

network. In the first version, which is shown in Fig. 5.2, thebiasing was done through the input and output signal pads us-ing external bias-tees, while in the second version the circuitis biased through separated bias pads. The measurement re-sults illustrated in the next section refer to the version withthe on-chip bias network.

5.1.3 Measurement Results

After fabrication, the S-parameters and the noise figure of thetraveling-wave amplifier were measured. Both measurementswere performed on-wafer using a Cascade Microtech probe-station together with an HP 8510B Network Analyzer and anHP 8970B Noise Figure Meter. For both S-parameter and noisefigure measurements the amplifier HEMTs were biased in thesaturation region ( (Vgs,Vds) = (0.15 V, 1.5 V) ). The totalcurrent drawn by the circuit is 90 mA and the total dissipatedpower is 135 mW for the version with external biasing, and540 mW for the version with on-chip biasing. This strong dif-ference between the two values is due to the fact that in thesecond case the bias current has to flow through the outputline termination resistance, which alone dissipates 75% of thetotal power. S-parameter measurements are shown in Fig. 5.3.A power gain between 10 and 11 dB is obtained up to 20 GHz.At low frequencies, the amplifier presents a gain peak which isdue to the limited size of RF-short on-chip capacitors and tothe strong frequency dependence of the line bend losses, whichwere not taken into account in the simulations and, therefore,not compensated [76]. Input and output return losses are ex-cellent up to 20 GHz, while the reverse isolation, which is notshown, remains below −30 dB in the entire measurement fre-quency range. In Fig. 5.4 the measured amplifier noise figureNF and S21 are reported as a function of frequency. In theband of interest, NF varies between 4.5 and 6 dB, reaching apeak of 8.5 dB at the lower and upper ends of the band. Thisrepresents a fairly good result for a traveling-wave amplifier,considering that the already poor low-noise characteristics of

76 MFOCS Circuits

0 5 10 15 20 25 30 −40

−30

−20

−10

0

10

20

Frequency (GHz)

S11

, S22

, S21

(dB

)

Fig. 5.3 Measured S21 (solid line), S22 (circles), and S11 (triangles) of the 0−20

GHz traveling-wave amplifer. The transistors bias point is (Vgs,Vds) = (0.15 V,

1.5 V), the total drain current is 90 mA.

0 5 10 15 20 254

5

6

7

8

9

10

11

12

13

14

Frequency (GHz)

S21

, NF

(dB

)

Fig. 5.4 Measured noise figure (squares) and S21 (solid line) of the traveling

wave amplifier. The transistors bias point is (Vgs,Vds) = (0.15 V, 1.5 V), the

total drain current is 90 mA.

16 GHz Up-converting Image-Rejection Resistive Mixer 77

this kind of circuit were worsen by the stabilization resistanceRs and by the mismatch of the termination impedances.The presented measurement results validate the used designapproach and confirm that the fabricated traveling-wave am-plifier is suitable, in terms of both gain, isolation, return losses,and noise figure to be employed in the MFOCS system.

5.2 16 GHz Up-converting Image-Rejection Resistive Mixer

5.2.1 Introduction

This up-converting resistive mixer uses a 16 GHz LO signalto perform the first up-conversion step in the system, i.e. toshift the data signal from the lower GHz range (1 − 2 GHz)up to the 14 − 15 GHz band. This up-conversion step is verycritical for the overall system performance. To avoid annoyingintermodulation problems, the mixer should ensure a good LOsuppression. An acceptable image-rejection is also necessary, inorder to keep the V-band mixer design simple and not to dealwith complicate filtering at 60 GHz.Because of the frequency spacing between the wanted and theunwanted signals, it would be very difficult to suppress thelocal oscillator and the upper side-band with filters; therefore,two resistive mixers in a singly balanced configuration wereemployed. The balanced configuration allows to suppress theupper side-band by means of phase-shifting at the IF and RFports. Additionally, resistive mixer cells have inherent stability,no DC power dissipation and guarantee a good LO suppressionthrough the HEMT itself without the need for a LO suppressionnetwork [77].


A schematic representation of the circuit is shown in Fig. 5.5.The mixer uses a single-balanced image-rejection configurationthat suppresses the upper-sideband (USB) signal. This kind ofconfiguration does not provide any performance improvement,i.e. the power of the USB is not recycled and used to improve

78 MFOCS Circuits

Fig. 5.5 Schematic of the image-rejection resistive mixer.

conversion gain.Two in-quadrature IF signals in the 1− 2 GHz band are givenas input to two identical single-transistor resistive mixers. The16 GHz LO signals are applied in phase, while the RF portsof the single mixers are connected to a 90o hybrid. The USBsignals combine at one of the hybrid output-ports with a 180o

phase difference, while the lower-sideband (LSB) signals at thesame port add in phase. In this way, USB rejection is achieved.The other port is terminated with a 50 Ω resistor.Figure 5.6 shows the schematic of the single-transistor resistive-mixer cell. The device is a 2 × 75 µm InP HEMT with a gatelength of 0.2 µm, presented in the previous chapters. By meansof large-signal simulations [78], the gate-source voltage for min-imum conversion losses (the drain-source voltage is 0 V) wasfound to be Vgs = −0.3 V, i.e. slightly above the threshold volt-age Vt of the device. The LO signal is applied without matchingto the device gate and modulates its drain-source resistance,which is responsible for the mixing effect. The IF and RF sig-nals are applied at the drain terminal, where LC matching net-works provide low return losses at both ports and good isola-tion between them. The RF signals of the two single-mixers arethen combined with a 16 GHz quadrature hybrid. A branch-


line coupler was chosen because of its straightforward imple-mentation. In order to avoid using long and space-consuming90o transmission lines at 16 GHz, the coupler was realizedusing capacitively-loaded high-impedance coplanar-waveguides(CPW). In this way, it was possible to reduce the circuit sizefrom the 4 mm2, that it would have occupied if regular 90o

transmission lines had been used, to 0.51 mm2 (87% size re-duction). The circuit is externally biased by means of bias-tees.In order to allow a better evaluation of the design approach,both the image-rejection and the single mixer were fabricated.A photograph of the fabricated image-rejection mixer is shownin Fig. 5.7.


After fabrication, both the resistive-mixer cell and the image-rejection mixer were measured. Characterization was performedon-wafer using a Cascade Microtech probe station, a HP 8510BNetwork Analyzer, and a HP 8565E Spectrum Analyzer. Themeasured S-parameters of the resistive-mixer cell are shown inFig. 5.8. Return losses at both IF and RF ports are well below−10 dB at the frequencies of interest and the IF-RF isolation isabove 20 dB at both IF and RF frequencies. For the resistive-mixer cell a peak conversion gain of −7 dB for 0 dBm LO power

Fig. 5.6 Schematic of the resistive-mixer cell.

80 MFOCS Circuits

Fig. 5.7 Photograph of the fabricated image-rejection resistive mixer. Chip size

is 2.25 × 2 mm2.

and 20 dB LO suppression were measured. This circuit does notprovide any image-rejection.After testing the resistive-mixer cell, the image-rejection mixer

was characterized. Measured conversion gain versus LO poweris plotted in Fig. 5.9. The peak value is −7.3 dB for an LOpower of 2 dBm. Considering that the LO port is not matched,this represents a very good result in terms of LO power require-ment. Matching the HEMT gate at 16 GHz, in fact, wouldmean using complicated LC network, since the device inputimpedance at that frequency is almost purely capacitive.Figure 5.10 shows the measured spectrum at the RF port whenthe IF signal is swept from 1 to 2 GHz and a 16 GHz LO signalwith a power of 0 dBm is applied. For the image-rejection mixerthe LO suppression is about 20 dB, while the USB rejectionvaries between 25 and 17 dB. The quite strong frequency depen-dence of image-rejection is due to the fact that the branch-linecoupler has a center frequency of 16 GHz, i.e. not in the middle


0 2 4 6 8 10 12 14 16 18 20−40

−35

−30

−25

−20

−15

−10

−5

0

5

Frequency (GHz)

S−

para

met

ers

(dB

)

Fig. 5.8 Measured IF-port return loss (solid line), RF-port return loss

(dot-dashed line), and RF-IF isolation (dashed line) of a single resistive mixer

cell. The HEMT bias point is (Vgs,Vds) = (-0.3 V, 0 V).

−14 −12 −10 −8 −6 −4 −2 0 2 4−22

−20

−18

−16

−14

−12

−10

−8

−6

PLO

(dBm)

Con

vers

ion

Gai

n (d

B)

Fig. 5.9 Measured conversion gain of the image-rejection mixer as a function

of the LO power. The HEMTs are biased at (Vgs,Vds) = (-0.3 V, 0 V).

82 MFOCS Circuits

of the USB signal band. This represents the best compromisebetween choosing a center frequency of 14.5 GHz, thus havingthe two LSB signals adding in phase at the RF port (best con-version gain) or choosing a center frequency of 17.5 GHz, thushaving a more precise 180o phase-shift of the USB signals (bestimage-rejection). Measurements of both the resistive-mixer celland of the image-rejection mixer confirm the validity of theused design approach and are in line with the best reportedresults for resistive-mixers in terms of conversion gain [79, 80],LO power requirement [81] and image-rejection [82]. The rea-sons of this performance are the careful design of the IF andRF matching networks and the excellent mixing characteris-tics of the used InP HEMTs at these frequencies. Measurementresults also show that, with its conversion gain, LO suppres-sion and USB rejection characteristics, the mixer meets all therequirements for the first up-conversion step of the MFOCSsystem.

13 14 15 16 17 18 19−45

−40

−35

−30

−25

−20

−15

Frequency (GHz)

PIF

(dB

m)

Fig. 5.10 Measured spectrum at the RF port of the image-rejection mixer. IF

frequency varies between 1 and 2 GHz, LO frequency is 16 GHz, IF power is

−10 dBm, LO power is 0 dBm.

16-48 GHz Active Frequency Triplers 83

5.3 16-48 GHz Active Frequency Triplers

5.3.1 Introduction

The frequency multiplier discussed in this sections is neededto generate a 48 GHz local oscillator from the 16 GHz refer-ence signal sent from the base station through the optical fiber.The generated mm-wave LO signal is used to drive the V-bandmixer which performs the final up-conversion step. In order tokeep the input power needed at 16 GHz as low as possible, thisfrequency tripler must have a good third harmonic conversiongain. Additionally, its output power must be high enough todrive the V-band mixer, i.e. at least −3 dBm, as next sectionwill show. Moreover, its output signal must have a good spec-tral purity, and the phase noise power degradation introducedin the frequency multiplication process should be kept as lowas possible. For this purpose, two frequency multipliers weredesigned and fabricated: a single-stage tripler and a tripler fol-lowed by a 48 GHz buffer amplifier.This part of the chapter is organized as follows: the next sectionpresents a theoretical analysis of HEMT single-device frequencymultipliers, which explains how frequency multiplication is ob-tained in active devices and presents a novel, simple approachto model this mechanism. In the other sections, the design ap-proach used for the proposed circuits and their measurementresults are presented.

5.3.2 Theoretical Analysis

Many theoretical analysis of FET frequency multipliers have al-ready been published, but none of them specifically addressesHEMT multipliers. Their common goal is finding an analyticalexpression for the device drain current ID that allows an a pri-ori determination of the optimal bias point, given the devicecharacteristics and the wanted output harmonic. Their com-mon assumption is that the drain current ID is generated by anideal current source solely controlled by the gate-source voltageVgs of the device. This represents a reasonable approximation

84 MFOCS Circuits

Fig. 5.11 Device waveforms for a generic device multipliers as described by the

Generalized Theory.

of how a FET-like device behaves when biased in the satu-ration region. Maas [83] represents the FET current using alinear transfer characteristic which clips to zero in the pinchoffregion. His approach was refined by O’Ciardha in a work de-scribed in [84], known as Generalized Theory (GT). The GTassumes that ID is a linear function of Vgs and clips to zerobelow the threshold voltage Vt of the device as well, but in-troduces a new clipping (to a maximum current Imax) whichoccurs when Vgs is higher that the gate-diode forward conduc-tion voltage Vf . For a sinusoidal excitation the drain currentcan be represented by the waveform shown in Fig. 5.11. Per-forming a Fourier analysis on this waveform to calculate theamplitude of the current harmonics results in the contour plotsof Fig. 5.12, where the amplitude of the first, second, and thirdharmonic is shown as a function of the gate-source bias pointVDC and of the amplitude of the sinusoidal excitation VAC .The agreement between the current harmonics simulated us-

ing the above described analysis and the measurement resultswas not found satisfactory. Therefore, a refinement of the Gen-


0.5 1 1.5−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3Second Harmonic

0.5 1 1.5−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3Third Harmonic

0.5 1 1.5−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3First Harmonic

VAC

(V)

VD

C (

V)

VAC

(V) VAC

(V)

Fig. 5.12 Contour plots of the amplitude of the first, second, and third har-

monic calculated with the Generalized Theory. Vt = −0.75 V, Vf = 0.9 V. White

areas corresponds to a larger amplitude.

eralized Theory was elaborated, which is presented in the re-maining part of this section. In a FET-like device the transcon-ductance gm is usually calculated as the first derivative of ID

versus Vgs. Therefore, the transconductance associated to thegeneric device described by the GT is a rectangular function ofVgs, i.e. zero below Vt and above Vf and constant between thesetwo voltages. Fig. 5.13 shows the measured gm(Vgs) of an InPHEMT fabricated with the process described in Chapter 3. Itcan be easily seen that the curve does not have a rectangularshape but, being bell-shaped, looks more like a triangle. Forthis reason, it was decided to approximate gm with a triangu-lar function of Vgs (Eq. (5.1)), integrate it toward Vgs to obtainan expression for the drain current (Eq. (5.2)) and use the (Eq.(5.2)) to calculate the current harmonics:

86 MFOCS Circuits

−1.5 −1 −0.5 0 0.5 1 1.50

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Vgs

(V)

g m (

S)

Fig. 5.13 Measured transconductance of an InP HEMT fabricated using our

in-house process. Gate length is 0.2 µm, gate width is 150 µm.

gm(Vgs) =

0, Vgs ≤ Vt

gmMAX

(Vgmax−Vt)

(

Vgs − Vt

)

, Vt ≤ Vgs ≤ Vgmax

gmMAX

(Vf−Vgmax)

(

Vf − Vgs

)

, Vgmax ≤ Vgs ≤ Vf

0, Vgs ≥ Vf

,

(5.1)

ID(Vgs) =

0, Vgs ≤ Vt

gmMAX

2(Vgmax−Vt)

(

Vgs − Vt

)2

, Vt ≤ Vgs ≤ Vgmax

gmMAX(Vgmax−Vt)

2+ gmMAX

2(Vf−Vgmax)(

V 2gmax + 2Vf (Vgs − Vgmax) − V 2

gs

)

, Vgmax ≤ Vgs ≤ Vf

gmMAX(Vf−Vt)

2, Vgs ≥ Vf

.

(5.2)This might look like trying to do one plus one in the hard

way, but it is not, since in this way ID can be calculated using alimited set of parameters which can be easily determined fromthe gm(Vgs) curve. These parameters are the threshold volt-age Vt, Vf (present also in the GT), the voltage Vgmax where gm


Fig. 5.14 (a) Triangular approximation of the transconductance of an InP

HEMT (solid line) compared with the transconductance assumed by the GT

(dashed line). (b) Device drain currents corresponding to a triangular approxi-

mation (solid line) and to the GT assumptions (dashed line).

reaches its peak, and the peak of the transconductance gmMAX .Fig. 5.14 shows the used trasnconductance function and itscorresponding ID compared to those of the GT. It is impor-tant to notice that the clipping effects are not anymore, likein the GT, the only source of non-linearity, since ID is now aquadratic function of Vgs. We will see that this has an effect onthe current harmonics prediction capability. Equation (5.2) isused to calculate the current harmonics when Vgs is a sinusoidalfunction of time. The contour plots resulting from the Fourieranalysis are shown in Fig. 5.15. These results might look likethose obtained with the GT (Fig. 5.12), but there is a signif-icant difference in the prediction of the optimal bias point forthe first and second harmonic, while for the third harmonic thedifference is not so marked. According to the prediction of theGT the best bias point for a frequency doubler is Vgs = −0.55V, while in our simulations it lies around Vgs = −0.4 V. This

88 MFOCS Circuits

0.5 1 1.5−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3Second Harmonic

VAC

(V)0.5 1 1.5

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3Third Harmonic

VAC

(V)0.5 1 1.5

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

0.3First Harmonic

VAC

(V)

VD

C (

V)


monic calculated using a triangular gm approximation. Vt = −0.75 V, Vf = 0.9

V, Vgmax = 0.1 V. White areas corresponds to a large amplitude.

0.1 0.5 1−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

Second Harmonic

VAC

(V)0.1 0.5 1

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

Third Harmonic

VAC

(V)0.1 0.5 1

−0.5

−0.4

−0.3

−0.2

−0.1

0

0.1

0.2

First Harmonic

VAC

(V)

VD

C (

V)


monic measured at a fundamental frequency of 16 GHz. White areas corresponds

to a large amplitude. Note that the VAC scale is different from the simulated

contour plots.


gives a difference of about 20% in the aperture, i.e., the frac-tion of the period T for which the drain current ID is actuallyflowing. For a tripler both methods predict an optimal biaspoint of Vgs = 0 V. To check the validity of this modeling ap-proach non-linear measurements were performed on the sameInP HEMT whose transconductance is shown in Fig. 5.13. Weapplied a fundamental tone at 16 GHz at the gate side of theHEMT, swept both its DC and AC value and measured thepower of the generated first, second and third harmonic at 16,32, and 48 GHz, respectively. The resulting contour plots areshown in Fig. 5.16. The agreement between the measurementsand the predictions of the “refined” theory (5.15)is excellentfor the first as well as for the second and the third harmonic.From both measurement results and simulations, it can beevinced that the optimal bias point for doubler operation isclass AB, while the for tripler operation is in class A, i.e. withthe device biased at the peak of the transconductance. This isunderstandable, if one thinks that at this bias point, for a suffi-ciently big input signal the output signal can be approximatedwith a square wave, which has a very high third harmonic con-tent. Apart from class A, a local maximum for third harmonicgeneration lies slightly above the threshold voltage Vt of thedevice.


The schematic representations of the two designed frequencytriplers are shown in Fig. 5.17 and 5.18. To obtain frequencymultiplication, the two circuits make use of the same designphilosophy, but differ for the output buffer stage present in thedouble-stage tripler. This section presents the design approachused for the frequency multiplier stage, followed by a brief de-scription of the 48 GHz buffer.It was shown in the previous section how for InP HEMTs themaximum third harmonic conversion gain is reached when thedevice is biased for class A operation. This corresponds to thesaturation region for the drain-source voltage Vds, and to the

90 MFOCS Circuits

Fig. 5.17 Schematic of the frequency tripler.

peak of the transconductance gm for the gate-source voltageVgs. However, this choice has the disadvantage of low conver-sion efficiency and leaves the designer with stability problemswhich are tedious to solve. Moreover, the large current flowinginto the device gate might lead to reliability problems. For thesereasons, in the design of the frequency multiplying stage, it wasdecided to set the device Vgs close to the threshold voltage Vt

and Vds in the saturation region. This bias point corresponds toa local maximum of the third harmonic conversion gain whichensures stability and higher conversion efficiency, but lower con-version gain.Since at 16 GHz the input impedance of the device (2 × 75µm InP HEMT) is almost purely capacitive, its input was not

Fig. 5.18 Schematic of the frequency tripler number 2.


Fig. 5.19 Photograph of the fabricated single-stage frequency tripler. Chip size

is 1.49 × 1.3 mm2.

matched at this frequency. The network at the gate side of theHEMT solely consists in an open circuit reflector stub with anelectrical length of 90o at 48 GHz. The stub reflects all the scat-tered power at 48 GHz back to the transistor thus increasingthe conversion gain, as it has already been shown in [86] forfrequency doublers. The distance between the reflector and thedevice was optimized using the look-up table based large signalmodel presented in the previous chapter and the Agilent ADSharmonic-balance simulator.The output network of the tripler consists of a 16 GHz resonatorwhich ensures the suppression of the fundamental frequency, anopen circuit stub which suppresses the second harmonic, anda matching network which provides an optimum large-signalmatching to 50 Ω. The biasing of the device gate is obtainedthrough a 10 KΩ resistor directly attached to the gate line;since there is virtually no current flowing into this terminal, thevoltage drop over the biasing resistor is negligible. The device

92 MFOCS Circuits

Fig. 5.20 Photograph of the fabricated double-stage frequency tripler. Chip

size is 3.15 × 1.39 mm2.

drain is biased through the output matching network using anon-chip RF-short capacitor. The buffer stage is a 48 GHz am-plifier with 50 Ω matching at the input and output port. Sincethe maximum output power of the multiplying stage does notexceed −10 dBm, the amplifier matching networks were opti-mized by means of small-signal simulations. Photographs of thefabricated circuits are shown in Fig. 5.19 and Fig. 5.20.


After fabrication, the frequency multiplication characteristicsof the two circuit were tested. Circuit characterization wasdone on-wafer using a Cascade Microtech probe station anda HP 8565E Spectrum Analyzer.Figure 5.21 shows the third harmonic conversion gain of asingle-stage tripler, while in Fig. 5.22 the third harmonic con-version efficiency is reported. The peak values for both conver-sion gain (−12 dB) and conversion efficiency (5.5%) are reachedfor an input power of about 1 dBm. The good conversion ef-ficiency is due to the fact that the DC power dissipation isextremely low (120 µW), as the HEMT is biased slightly abovethe threshold voltage Vt. Figure 5.23 shows the output power atthe fundamental frequency and at 48 GHz as a function of theinput power at 16 GHz for a single-stage tripler. The outputpower at 48 GHz does not saturate at a maximum value, but


reaches a peak of about 10 dBm and then drops down. This

−10 −8 −6 −4 −2 0 2 4 6 8−24

−22

−20

−18

−16

−14

−12

−10

Pin

(dBm)

3rd

Har

mon

ic C

onve

rsio

n G

ain

(dB

)

Fig. 5.21 Third harmonic conversion gain as a function of the input power for

a single-stage tripler. The device bias point is (Vgs, Vds) = (−0.375 V,1.2 V).

is in agreement with both theory and experimental results [84].As it can be seen from the graph the first harmonic suppres-sion is always better than 25 dB and the power ratio betweenthe third harmonic and the fundamental is 15 dB at 0 dBminput power. The power ratio between the third harmonic andthe second harmonic is also excellent, greater than 35 dB. Theactual output power at 32 GHz could not be measured becauseit was always below the sensitivity of the HP 8565E spectrumanalyzer used for the measurements. This is an indication of thegood spectral purity of the output signal which is in line withthe best reported GaAs PHEMTs [87] single-device frequencymultipliers. In Fig. 5.24 the third harmonic conversion gain ofthe double-stage tripler is presented. The peak value (−7.3 dB)is reached for an input power of 3 dBm. This means that thebuffer amplifier manages to provide 5 dB additional gain at 48GHz, which is quite low. A possible explanation for the missing

94 MFOCS Circuits

gain can be the mismatch between the two stages. The out-put power harmonics of the double-stage tripler are shown inFig. 5.25. The measured maximum output power is −2.5 dBm,reached for an input power of 3 dBm. The fundamental sup-pression is around 30 dB, while the power ratio between thethird harmonic and the fundamental is 20 dBc at the peak ofthe output power. The power ratio between the third and thesecond harmonic is around 37 dB. The double-stage tripler,with respect to the single-stage one, shows significant improve-ments in terms of conversion gain, output power, and spectralpurity of the generated signal. The conversion efficiency dropsdown by an order of magnitude due to the higher DC powerdissipation.Phase noise is an important characteristic for frequency gener-ation systems and, therefore, for frequency multipliers. For anideal frequency multiplier, the phase noise power degradationis given by 20 · log(N), where N is the multiplication order. For

−10 −8 −6 −4 −2 0 2 4 6 80

1

2

3

4

5

6

Pin

(dBm)

3rd

Har

mon

ic C

onve

rsio

n E

ffici

ency

(%

)

Fig. 5.22 Measured third harmonic conversion efficiency as a function of the

input power for a single-stage tripler. The device bias point is (Vgs, Vds) =

(−0.375 V,1.2 V).


−10 −8 −6 −4 −2 0 2 4 6 8−45

−40

−35

−30

−25

−20

−15

−10

Pin

(dBm)

Out

put P

ower

(dB

m)

Fig. 5.23 Measured output power at 48 GHz (circles) and at 16 GHz (squares)

as a function of the input power for a single-stage tripler. The device bias point

is (Vgs, Vds) = (−0.375 V,1.2 V).

−10 −8 −6 −4 −2 0 2 4 6−26

−24

−22

−20

−18

−16

−14

−12

−10

−8

−6

Pin

(dBm)

3rd

Har

mon

ic C

onve

rsio

n G

ain

(dB

)

Fig. 5.24 Measured third harmonic conversion gain as a function of the input

power for a double-stage tripler.

96 MFOCS Circuits

−10 −8 −6 −4 −2 0 2 4 6−60

−50

−40

−30

−20

−10

0

Pin

(dBm)

Pou

t (dB

m)

Fig. 5.25 Measured output power at 48 GHz (circles), at 32 GHz (triangles),

and at 16 GHz (squares) as a function of the input power for a double-stage

tripler.

a frequency tripler, this corresponds to a minimum theoreticalvalue of 9.5 dB [88]. The measured phase noise degradation asa function of the frequency offset from carrier for the single-stage tripler is reported in Fig. 5.26. It was measured at thepeak of the conversion gain using two different signal gener-ators. It can be seen that all the values are spread around 9dB, except for one point which surely is the result of a mea-surement error. These results are in line with already publishedresults [89], but their discussion is entangled by the difficultyto estimate their accuracy. The value of ± 1 dB accuracy foundin the literature for analogous measurements, in fact, can notbe supported by any experimental results.To the author’s knowledge these are the first InP HEMT fre-quency triplers ever reported [90].

V-band Up-Converting Active Mixer 97

10 Hz 100 Hz 1 KHz 10 KHz 100 KHz 1 MHz5

6

7

8

9

10

11

12

13

Frequency Offset

Pha

se N

oise

Deg

rada

tion

(dB

)

Fig. 5.26 Measured phase noise power degradation as a function of the fre-

quency offset from the carrier for a single-stage tripler. The device bias point

is (Vgs, Vds) = (−0.375 V,1.2 V), input power is 1 dBm. Two different signal

generators have been used: HP 83650L (circles) and Rohde & Scwharz SMP04

(squares).

5.4 V-band Up-Converting Active Mixer

5.4.1 Introduction

The active mixer described in this section converts the datasignal in the 14 − 15 GHz band up to the V-band (62 − 63GHz) using the 48 GHz local oscillator signal generated by thefrequency tripler discussed in the previous section. The circuitdesign was optimized to minimize the LO power requirementswhile keeping conversion losses as low as possible. The useddesign approach and the most significant measurement resultsare presented.


A schematic representation of the designed up-converting mixeris shown in Fig. 5.27. The circuit is a single-device mixer whichuses a two-finger InP HEMT with a gate length of 0.2 µm anda total gate width of 150 µm. A gate injection topology was

98 MFOCS Circuits

preferred because it has shown the best conversion gain forlow LO power levels [77, 91–95]. Because of the phase modu-lation scheme used in the system, the modest linearity of thiskind of mixers does not represent a problem. The device biaspoint was chosen to minimize the variation of the drain-sourceconductance |gds| and to maximize the fundamental frequencycomponent of the transconductance |gm|. This corresponds toa gate-source voltage Vgs slightly above the threshold voltageVT of the device and to a drain-source voltage Vds in the satu-ration region. While optimizing the circuit for low LO power,it was found that the optimum large-signal matching networksboth at the gate and at the drain side of the transistor canbe very well approximated with matching networks calculatedusing only small-signal simulations of the device, which are lesstime-consuming to optimize in the used circuit simulator. Thenetwork at the gate side of the mixer was designed to provide50 Ω matching at both IF (16 GHz) and LO (48 GHz) frequen-cies. Since the two signals must be applied at the same port,the use of an external diplexer is necessary. The gate bias net-work also includes a small (5 Ω) resistor, which is needed toensure unconditional stability to the circuit. This resistor hasnegligible effects on the IF signal up-conversion. The networkat the drain side of the device ensures the suppression of bothLO and IF signals and provides 50 Ω matching at RF frequency

Fig. 5.27 Circuit schematic of the designed up-converting mixer.


Fig. 5.28 Photograph of the fabricated V-band up-converting mixer. Chip size

is 3.7 × 1.4 mm2.

(64 GHz). Obtaining a good LO suppression is important be-cause it allows the HEMT to remain in the saturation regionover the entire LO cycle thus keeping |gds| as constant as possi-ble. This is achieved by an open circuit stub with an electricallength of 90o at 48 GHz. The suppression of the IF signal is ob-tained using an LC resonator instead of a stub in order to keepthe circuit size small. Since no models were available in the cir-cuit simulator (Agilent ADS), coplanar discontinuities such ascross-junctions, T-junctions, open circuits, and short circuitswere modeled using the SONNET electromagnetic simulator,using the procedure described in Chapter 3. A photograph ofthe mixer is shown in Fig. 5.28. Chip size is 3.7 × 1.4 mm2.


After fabrication, the characteristics of the up-converting mixerwere measured on-wafer using a Cascade Microtech probe sta-tion, an HP 8510C Network Analyzer and an HP 8970B Spec-trum Analyzer. Measured small-signal return losses at the in-put and output ports of the mixer are shown in Fig. 5.29. 50 Ωmatching is excellent both for the IF signal (16 GHz) and theLO signal(48 GHz) as well as at RF frequency (64 GHz). Af-ter measuring the S-parameter, the non-linear characteristics ofthe circuit were tested. In order to combine LO and IF signal atthe gate port a Test Port Combiner from HP was used. In Fig.5.30, the mixer conversion gain at 64 GHz is plotted versus theLO power. For an LO power of −10 dBm the conversion gain is

100 MFOCS Circuits

0 10 20 30 40 50 60 70 80−30

−25

−20

−15

−10

−5

0

Frequency (GHz)

S11

, S22

(dB

)

Fig. 5.29 Measured S11 (solid line, IF and LO port) and S22 (dashed line, RF

port) of the up-converting mixer. The device bias point is (Vgs,Vds)=(−0.3 V,

1.2 V).

−3 dB and keeps increasing with increasing LO power until itreaches its peak value for a LO power of 0 dBm. In a LO power

−10 −8 −6 −4 −2 0 2 4−5

−4

−3

−2

−1

0

1

LO Power (dBm)

Con

vers

ion

Gai

n (d

B)

Fig. 5.30 Measured conversion gain versus LO power. The device bias point is

(Vgs,Vds)=(−0.3 V, 1.2 V), IF frequency is 16 GHz, IF power is −30 dBm.


61 62 63 64 65 66−10

−8

−6

−4

−2

0

2

Frequency (GHz)

Con

vers

ion

Gai

n (d

B)

Fig. 5.31 Measured conversion gain as a function of the RF frequency. The

device bias point is (Vgs,Vds)=(−0.3 V, 1.2 V), LO power is −1.7 dBm, IF

power is −30 dBm.

−25 −20 −15 −10 −5 0−60

−50

−40

−30

−20

−10

0

Pin

(dBm)

Pou

t (dB

m)

f1

2f1−f

2

Fig. 5.32 Measured third order intercept point of the up-converting mixer.

The device bias point is (Vgs,Vds)=(−0.3 V, 1.2 V), LO power is −3 dBm,

f1 = 15.9 GHz, f2 = 15.901 GHz.

102 MFOCS Circuits

range between −4 and 0 dBm, conversion gain remains almostconstant, showing a variation of only 0.5 dB. This validatesthe low-power design approach described in the previous sec-tion. For an LO power higher than 0 dBm, the conversion gaindrops dramatically. This is due to the fact that, for a greaterLO power, acceptable return losses at IF and RF frequenciesare not guaranteed anymore, since the matching is optimizedfor small-signal behavior. LO suppression was also measuredand was found to be higher than 30 dB at 48 GHz. Measuredconversion gain versus RF frequency is shown in Fig. 5.31 fora LO power of −1.7 dBm and a IF power of −30 dBm. Thepeak conversion gain is 1 dB at 64.5 GHz. The strong frequencydependence of conversion gain is due to narrow-band behaviorof the input matching at IF frequency. In order to evaluatethe linearity of the up-converting mixer, the 1 dB compressionpoint and the third order intercept point (IP3) were measured.This was done at a IF frequency of 15.9 GHz with −3 dBmLO power. The 1 dB compression point was reached for an IFpower of −10 dBm. The IP3 was found to be −1.6 dBm, as canbe seen from Fig. 5.32. Since the maximum expected IF poweris −15 dbm, the rather low compression point and IP3 allowthis mixer to be used in the MFOCS system. To the author’sknowledge, the measurement results, in terms of both LO powerrequirements and conversion gain, are the best ever reported fora single-transistor up-converting mixer in the V-band [96].

6

Flip-Chip Bonding

From their measurement results, it is evident that the circuitspresented in the previous chapter can be successfully employedin the fabrication of the millimeter-wave fiber-optic commu-nication system proposed in chapter 2. For the system to beassembled, an appropriate mounting technique needs to be cho-sen, optimized, and tested on the MMICs of the available chipset.When mounting different integrated circuits on a board, mi-crowave designers can choose between two possible solutions:wire bonding or flip-chip bonding. The signal frequency in thetransitions between the different components of the proposedsystem varies from the lower GHz range to the V-band. Theneeded mounting technique should, therefore, be easy to handleat very low frequencies as well as at 64 GHz. It is not controver-sial to say that wire bonding does not meet this requirement;although this technique can be successfully used for mountingmillimeter-wave MMICs, its high sensitivity to mechanical tol-erances makes it arduous to employ for V-band applications.For these reasons, flip-chip bonding was investigated [97,98].This chapter illustrates how a flip-chip bonding technique wasimplemented and optimized for applications up to the V-band,and which kind of problems arose while testing it on the fabri-cated InP MMICs. In the first part of the chapter, some generalinformation about flip-chip bonding its given and the advan-tages of this technique are presented. In the second section, theused flip-chip process and its optimization for V-band appli-

104 Flip-Chip Bonding

cation are described. The chapter ends with the depiction ofthe encountered problems and with the explanation of why noassembled system is presented in this dissertation.

6.1 General Aspects

Flip-chip bonding is the direct electrical connection of face-down chips onto a substrate or a circuit board by means ofconductive bumps placed on the chip bond-pads. This is incontrast with wire bonding technology, which uses face-up chipswith a wire connection to each pad. The bumps guarantee boththe electrical and the mechanical connection between the chipand the mounting substrate, while with wire bonding the me-chanical connection is ensured by gluing or soldering the chipdirectly on the board.Flip-chip interconnections were introduced for the first timeby IBM in 1963 in the assembling process of mainframe com-puter boards. At present, they are manly used for automotiveapplications, electronic watches, and an increasing percentageof cellular phones, high speed microprocessors, and microwavecircuits. Nowadays, more than 600.000 components are annu-ally produced using flip-chip technology, with a growth rate ofalmost 50% per year. This means that, within few years, 10%of the fabricated semiconductors chips will be assembled usingflip-chip bonding technology [99,100].

6.2 Why Flip-Chip

For integrated circuits in general and microwave componentsand circuits in particular, there are numerous benefits in em-ploying flip-chip bonding instead of wire bonding. The mostimportant advantages are listed below.

• Superior Electrical Performance. Compared to bondwires, the interconnections between a flip-chip mounted cir-cuit and the substrate have a much shorter length. Typ-ical values are in the order of 30 − 50 µm for a bump

Process description 105

and 500 − 1000 µm for a bond wire. Flip-chip intercon-nections have, therefore, lower parasitic inductance, capac-itance, and resistance, and benefit of minimal RF losses.Additional increase in the bond wires length can result bymounting together substrates of different heights, or by thepresence of contact pads in the central part of the chip.This problems can be easily circumvented with flip-chip.

• Minimal Mounting Area. With this technique the mount-ing area on the substrate can be reduced. Both the bias andthe signal lines, in fact, can be drawn below the chip anddo not have to stop at its borders. This has positive effectson the package outline as well.

• Higher Reproducibility and Lower Costs. With flip-chip bonding, the placement of the chips is fast, easy tohandle, and more precise than with wire bonding. The in-creased precision has as a consequence the elimination oftedious tuning steps, sometimes necessary with wire bond-ing techniques, especially at V-band frequencies. Increasingthe mounting speed and removing the tuning steps can leadto important cost reductions.

The circuits designed for the MFOCS project, were fabricatedusing a coplanar wave-guide technology. To take full advan-tage of this fact, it was decided to mount them on a coplanarwave-guide environment. It has been shown that, for flip-chipbonding, it is the combination CPW on CPW which resultsin the best RF transition between the mounting substrate andthe chip. Transition return losses lower than −20 dB up tothe W-band have already been demonstrated [101], which havenot been obtained with other combinations, like e.g. CPW onmicrostrip lines. Moreover, using coplanar wave-guides on themounting board has the additional advantage of requiring noback-side metallization and via holes.In the following section the flip-chip bonding technique used atthe Microwave Electronics Laboratory will be illustrated andits optimization for V-band applications explained.


6.3 Process description

A schematic representation of the flip-chip bonding techniqueused for this work is shown in Fig. 6.1. The first step (Fig. 6.1 (a))is the bump formations using the ball bump method. By meansof a wire bonding tool a “ball” of gold is deposited over a 10 milthick ceramic (Al2O3) substrate. Ceramic was chosen becauseof its mechanical characteristics; its 10 mil thickness becauseof its suitability for V-band coplanar wave-guide applications.A SEM picture of a deposited gold bump is shown in Fig. 6.2.It has diameter is 73 µm and an overall height of 50 µm. Theacute tail is easy to deform under thermal compression and re-

Fig. 6.1 Schematic description of the used flip-chip bonding technique. (a)

Bumps formation with the “ball bump” method; (b) Chip mounting by means

of thermal compression; (c) Final result.

Optimization for V-band applications 107

Fig. 6.2 Scanning Electron Microscope picture of a “gold ball” bump. The

bump diameter is 73 µm.

sults in a very stable bonding area.After the bump formation, the chip and the substrate are bondedby means of thermal compression (Fig. 6.1 (b)). This is car-ried out using a Fineplacerr − 96 LAMBDA bonding systemfrom Finetech [102]. This system, thanks to an alignment sys-tem with a fixed optical beam splitter, allows flip-chip bondingwith 1 µm placement precision. The size of the bonded chip canvary from 100 µm to 35 mm. The substrate holder is heated at150 C, while the temperature of the bonding tool head, whichholds the chip, is ramped up to 300 C. Once the temperaturesof both the holder and the head are stable, the chip and thesubstrate are pressed together for 10 s. The force used in thisoperation determines the thickness of the air gap between chipand substrate, i.e. the final height of the bump (Fig. 6.1 (c)).

6.4 Optimization for V-band applications

Optimizing a flip-chip RF interconnection means to minimizeits insertion losses and to keep the reflections due to the signaltransition between the substrate and the mounted chip as small


as possible. A good figure of merit for the quality of a transi-tion is its return loss: return losses below −20 dB indicate analmost ideal transition.In order to minimize the parasitic effects due to the bumps, itis necessary to keep them as short as possible [103–106]. Un-fortunately, the bump height can not be arbitrarily reduced.When the distance between the on-chip CPWs and the mount-ing substrate become comparable with the ground-to-groundspacing of the CPW, the impedance of the coplanar wave-guidechanges. This is called impedance detuning due to proximityeffects. A trade-off between the bump height and the proximityeffects must be found. Figure 6.3 and 6.4 show the simulatedimpedance detuning of a 50 Ω CPW line on a ceramic sub-strate and on an InP substrate, respectively, as a function ofthe distance from a ceramic substrate. In both cases, for a bumpheight shorter than 25 µm, the impedance detuning becomessignificant and has to be taken into account in the circuit de-

0 25 50 75 100 125 150−12

−10

−8

−6

−4

−2

0

Bump Height (um)

Impe

danc

e D

etun

ing

(%)

Fig. 6.3 Simulated impedance detuning of a 50 Ω CPW line on a ceramic

substrate mounted on a ceramic substrate. The CPW ground-to-ground spacing

is 210 µm.


0 25 50 75 100 125 150−10

−5

0

5

Bump Height (um)

Impe

danc

e D

etun

ing

(%)

Fig. 6.4 Simulated impedance detuning of a 50 Ω CPW line on an InP substrate

mounted on a ceramic substrate. The CPW ground-to-ground spacing is 70 µm.

sign phase. This is not acceptable if the circuits to be mountedhave already been fabricated. Therefore, a 25 µm bump heightwas chosen as the best compromise between impedance detun-ing and a short bump height.In order to determine the necessary bonding force to obtain a25 µm bump height, bonding tests with different applied forceswere performed and the resulting bump heights measured us-ing a scanning electron microscope. The results are plotted inFig. 6.5. In order to achieve a 25 µm bump height a force perbump of 0.5 N is necessary. The most important parametersinvolved in the process are summarized in Table 6.1.Once the best compromise between bump height and impedance

detuning was found and the relevant process parameters wereoptimized, test structures were fabricated with the purpose ofevaluating and characterizing the signal transition between themounting substrate and the chip. The fabricated test struc-tures consist of a 50 Ω CPW on a 10 mil ceramic substratemounted on another 10 mil ceramic substrate in series with two


Table 6.1 Summary of the most important parameters of the flip-chip bonding

process used in this work.

Parameter Value

Substrate Temperature 150 CChip Temperature 300 C

Force/Bump 0.5 NCompression Duration 10 sResulting Bump Height 25 µm

Resulting Bump Diameter 90 µm

50 Ω CPWs, like in Fig. 6.1 (c). The measured S-parameter ofthe series connection of the three lines are shown in Fig. 6.6.They were measured using a Cascade Microtech probe stationand a HP 8510B Network Analyzer. The quality of the signaltransitions, both in terms of return loss and insertion loss, isacceptable up to 20 GHz, but it significantly deteriorates forfrequency approaching the millimeter-wave region.In order to use this flip-chip process for V-band applications,

further optimization is needed. Since all the possible process pa-

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.918

20

22

24

26

28

30

32

Force/Bump (N)

Bum

p H

eigh

t (um

)

Fig. 6.5 Bump height as a function of the bonding force per bump.


0 5 10 15 20 25−1.5

−1

−0.5

0

Frequency (GHz)

Inse

rtio

n Lo

ss (

dB)

0 5 10 15 20 25−60

−40

−20

0

Frequency (GHz)

Ret

urn

Loss

(dB

)

Fig. 6.6 Measured S-parameters of a three-lines 50 Ω test structure fabricated

to evaluate the quality of a CPW-to-CPW signal transition.

rameters have already been tuned, an additional improvementof the transition RF performance can be obtained only by us-ing matching networks. For this purpose, an equivalent circuitrepresentation of a flip-chip discontinuity is required [107–109].By fitting the measured S-parameter with different equivalentcircuits, it was found that a flip-chip discontinuity can be ef-fectively modeled by a lumped capacitor to ground. In thiscase, the value of the capacitance was calculated to be 20 fF.

Fig. 6.7 Equivalent circuit of a flip-chip CPW-to-CPW signal transition.


The equivalent circuit representation is depicted in Fig. 6.7.The capacitive behavior of this transition can be explained byconsidering the three bumps needed for a CPW-to-CPW con-nection as two parallel three-dimensional capacitors from thesignal pad to the two ground pads. Since the interconnectionis rather short compared to its diameter, no inductive behaviorcan be noticed and the capacitive effect dominates.There are two possible approaches to reduce the effects of thecapacitance associated with a flip-chip transition [101,110,111].The bump dimensions can be reduced in order to decrease theparasitic capacitance value. This means changing from a goldball deposition process to a process where the bumps are de-fined by optical lithography and then deposited by electroplat-ing. If this is not possible, the parasitic effects can be counter-balanced by means of a compensation network at each transi-tion. For sake of simplicity, the second solution was adopted.A compensation network consisting of a series connection of ahigh-impedance and of a low-impedance transmission line was

Fig. 6.8 Photograph of the fabricated compensation network.

Application to InP MMICs 113

0 20 40 60 80 100 120−2

−1.5

−1

−0.5

0

Frequency (GHz)

Inse

rtio

n Lo

ss (

dB)

0 20 40 60 80 100 120−50

−40

−30

−20

−10

0

Frequency (GHz)

Ret

urn

Loss

(dB

)

Fig. 6.9 Measured S-parameters of a three-lines 50 Ω test structure with com-

pensation networks at each transition.

designed by means of both circuit simulations, for a first ap-proximation, and full-wave electromagnetic simulations. A pic-ture of the fabricated compensation network is shown in Fig.6.8.A new test structure, equivalent to the one used to evaluate thenon-compensated transitions, but with compensation networksat each transition was fabricated and measured. The resultingS-parameters are shown in Fig. 6.9. Both the measured inser-tion loss and return loss are excellent up to 80 GHz. The stronglow-frequency variations of the measured insertion loss are dueto calibration problems. The measurement results confirm thatthis simple and low-cost flip-chip bonding technique achievesexcellent transition performance up to the V-band and beyond.

6.5 Application to InP MMICs

The flip-chip bonding technique described in the previous sec-tion was developed in order to mount the fabricated MFOCScircuits on a ceramic substrate, thus integrating the complete


Fig. 6.10 Photograph of the 14.5 GHz amplifier. Chip size is 2.09× 1.56 mm2.

system on a single board. Unfortunately, because of technol-ogy related problems described in the next section, it was notpossible to assemble the complete system. In this section, aflip-chip bonded 14.5 GHz monolithically integrated amplifieris presented as an example of how the developed flip-chip tech-nique can be successfully employed to mount InP MMICs on a10 mil ceramic substrate.The mounted circuit is a single-transistor amplifier, fabricatedwith the in-house process presented in chapter 3. Both the in-

Fig. 6.11 Photograph of the flip-chip mounted 14.5 GHz amplifier.

Encountered Problems 115

0 5 10 15 20 25 30 35 40 −30

−25

−20

−15

−10

−5

0

5

10

15

Frequency (GHz)

S−

para

met

ers

(dB

)

S11

S21

S22

S12

Fig. 6.12 Comparison between on-wafer measured S-parameters of the un-

mounted 14.5 GHz amplifier (solid line) and S-parameter measurements of the

flip-chip bonded circuit (dashed line).

put and the output ports are matched to 50 Ω by means ofsingle-stub matching networks. In order to achieve uncondi-tional stability, the HEMT source was inductively loaded withhigh impedance coplanar wave-guides. The amplifier is DC iso-lated at both ports and it is biased through separated bias pads.A photograph of the mounted circuit is shown in Fig. 6.10.On the mounting substrate, two access coplanar wave-guidesterminating with the compensation networks described in theprevious section were fabricated together with DC bias lines.Figure 6.11 shows a photograph of the flip-chip mounted am-plifier. The circuit was measured on Cascade Microtech probestation with an HP 8510B Network Analyzer, before and afterbeing mounted. A comparison between on-wafer measurementsof the unmounted amplifier (solid line) and the S-parametersof the flip-chip mounted circuit (dashed line) is presented inFig. 6.12. The very small difference between the two measure-ments up to 40 GHz confirms the good quality of the optimizedflip-chip transitions and validates the used bonding technique.


6.6 Encountered Problems

A major problem, encountered while mounting the fabricatedcircuits, represents the reason why no assembled system is pre-sented in this work. All the flip-chip bonded circuit, exceptthe one described in the previous section, did not stick tothe mounting substrate, but fell immediately off. From opticalmicroscope inspections, it was noticed that, after the detach-ment of the circuit, all the bumps still stuck to the mountingsubstrate and presented parts of the galvanic metallization at-tached to them. It was concluded that the galvanic metalliza-tion had a good adherence with the bonding bumps, but notto the lower metal layers of the integrated circuit.A review of the galvanic process used in the circuit fabricationis useful to understand this phenomenon. As shown in Fig. 3.4,the galvanic process consists of two steps; first, a 20 nm thickNickel layer is evaporated, then, the galvanic gold is grown byelectroplating on top of the Nickel. While removing the un-wanted galvanic metal, the Nickel layer below the gold mustalso be etched away. This etching process can cause the removalof the Nickel from underneath the wanted galvanic gold. If toomuch Nickel is removed (a minimal under-etching must alwaysbe expected), a lack of adherence of the galvanic metallizationwill be experienced, which seriously impacts any bonding pro-cess. A way to circumvent this problem is replacing the Nickellayer with a Titanium layer in the galvanization process.

6.7 Conclusions

A simple low-cost gold-ball flip-chip bonding process was ef-fectively optimized for V-band applications. For the optimizedCPW-CPW flip-chip transitions, return losses smaller than −20dB up to 80 GHz were measured. The application of the op-timized bonding technique to microwave amplifiers was alsodemonstrated. In order to reliably flip-chip bond the circuit fab-ricated with the process described in chapter 2, technological

Conclusions 117

problems related to the adherence of the galvanic metallizationstill need to be solved.

7

Summary, conclusions, andoutlook

The following sections contain a recapitulation of the main re-sults presented in this dissertation, describe the unsolved prob-lems, and indicate some directions for future work on the sub-ject.

7.1 System Design

The second chapter presents the different aspects that had tobe taken into account in the design of the MFOCS system andcontains a description of the proposed architecture. There aretwo main difficulties in the design of a mm-wave fiber-radio sys-tem. On one hand, the need for a stable millimeter-wave localoscillator signal in the hub can not be fulfilled by employingmillimeter-wave local oscillators, which do not guarantee therequired frequency stability. On the other hand, the transmis-sion losses caused by chromatic dispersion in the optical fibers,set an upper limit for the frequency of the transmitted opticalsignal. The chosen architecture circumvents these problems bygenerating, in the base station, a low-frequency (16 GHz) LOsignal which is added to the data signal and then transmittedthrough the fiber. By limiting to 16 GHz the maximum fre-quency of the transmitted optical signal, chromatic dispersioneffects are prevented and the bandwidth specifications of boththe laser driver and the optical receiver are relaxed. In the hub,the high-frequency local oscillator is obtained by frequency mul-tiplication of the low-frequency reference signal. The frequency

120 Summary, Conclusions, and Outlook

up-conversion of the data signal is performed in two succes-sive mixing steps. This is preferable to a direct up-conversionbecause it significantly eases the mixer design. The hub specifi-cations, in terms of photo receiver noise figure, minimum trans-mitted power etc., were derived by a careful investigation of theoptical link and of the radio link. Radiation constraints due tothe current regulation on electromagnetic emission were alsotaken into account.

7.2 InP HEMT Large-Signal Modeling

Chapter 4 describes an efficient procedure to extract a look-uptable-based InP HEMTs large-signal model which takes impactionization effects into account. The proposed large-signal modelwas verified up to 110 GHz with S-parameter measurements,and up to the V-band with non-linear measurements. The non-linear model performance in the impact ionization frequencyregion was also tested by means of third order intermodula-tion measurements. It is also demonstrated that the describedlarge-signal model can be employed to successfully predict theperformances of amplifiers and mixers up to the V-band. Thisis the first InP HEMT look-up table-based large-signal modelwhich gives a large-signal representation of impact ionization.

7.3 Circuit Design and Fabrication

Chapter 5 presents the most important MMICs designed in theframe of the MFOCS project. The circuits were simulated us-ing the large-signal model described in chapter 4 and fabricatedusing the InP HEMT in-house process of the Microwave Elec-tronics Laboratory.The 0− 20 GHz TWA amplifies the two components of the re-ceived optical signal, i.e. the data part and the 16 GHz carrier.It provides input and output return losses smaller than −20dB and a gain of more than 10 dB, from DC to 20 GHz. Themeasured noise figure is perfectly suitable for the applicationas an optical receiver amplifier.

Flip-Chip optimization for V-band Applications 121

The 16 GHz image-rejection resistive mixer performs the firstup-conversion step. Key feature of this components is a goodrejection of the upper side-band, which is needed to avoid com-plicated filtering in the V-band. The mixer has −7.3 dB con-version gain for 3 dBm LO power, an image rejection of over 17dB, and more than 20 dB LO suppression. Its measured per-formance fulfills all the specifications of the MFOCS system.The 16 − 48 GHz frequency tripler converts the low-frequencyreference signal sent to the hub from the base station to amillimeter-wave LO signal, which is used in the hub to per-form the second frequency up-conversion. The tripler, inte-grated with a 48 GHz buffer amplifier, shows a maximum out-put power of −2.5 dBm, over 30 dB suppression of the fun-damental frequency, and 37 dB power ratio between the thirdand the second harmonic. To the author’s knowledge, this isthe first InP MMIC frequency tripler ever presented.The V-band active mixer performs the final data signal up-conversion, from 16 GHz to the V-band, using the 48 GHz LOsignal generated by the frequency tripler. The fabricated cir-cuit shows a peak conversion gain of 1 dB at 64.5 GHz for −1.7dBm LO power, and over 30 dB suppression of the local oscilla-tor signal. To the author’s knowledge, the measurement results,in terms of both LO power requirements and conversion gain,are the best ever reported for a single-transistor up-convertingmixer in the V-band.

7.4 Flip-Chip optimization for V-band Applications

Chapter 6 describes the optimization of a gold-ball flip-chipbonding technique for V-band applications. First, the CPW-CPW flip-chip transition was carefully characterized and itslumped element representation determined. If was found thatthe parasitic effects in a flip-chip mounted CPW-CPW inter-face can be effectively represented by a capacitor to ground.Based on this equivalent circuit, a on-substrate coplanar wave-guide compensation network was designed which allows an ex-

122 Summary, Conclusions, and Outlook

cellent signal transition (return losses lower than 20 dB) up to80 GHz. This technique is the most promising way to assemblethe complete system together. Technological problems relatedto the adherence of the galvanic metallization layer still have tobe solved to effectively employ flip-chip bonding for the circuitsfabricated with the IFH in-house process.

7.5 Conclusions and Future Work

The work presented in this dissertation demonstrates that theInP HEMT process presented in chapter 3 can be successfullyemployed for the fabrication of a V-band fiber-radio communi-cation system. Thanks to an appropriate system design, possi-ble problems related to the generation of a stable millimeter-wave local oscillator signal and to the effects of chromatic dis-persion in the optical fibers were circumvented.By using accurate models for both active and passive compo-nents combined with a good process stability, the key mono-lithically integrated microwave circuits were designed and man-ufactured by the IFH clean room team.A flip-chip bonding technique was developed and optimized forV-band applications. The use of flip-chip technique instead ofwire-bonding eases the assembling of the complete system. Theonly encountered obstacle is related to the adherence of thebonding pads galvanic metallization. Solving it is a necessarystep for the realization of the complete system.Additional starting-points for future research work are offeredby this dissertation. The development of an empirical InP HEMTlarge-signal model with a larger operating range will allow amore extensive optimization of non-linear circuits operatingwith large input and output signals.A great simplification in the system architecture can be ex-pected from the employment of sub-harmonically pumped up-and down-converting mixers with acceptable conversion losses.Finally, the fabrication of InP HEMTs with the same high-frequency performance but with improved breakdown charac-

Conclusions and Future Work 123

teristics, will ease the design of the high-gain, high-output-power V-band amplifiers indispensable for this kind of systems.

Bibliography

[1] P. Smulders et al., “Exploiting the 60 GHz Band for LocalWireless Multimedia Access: Prospects and Future Di-rections”, IEEE Communication Magazine, pp.140-147,January 2002.

[2] L.M. Correia et al., “An Overview of Wireless BroadbandCommunications”, IEEE Communication Magazine, pp.28-33, January 1997.

[3] M. Chelouche et al., “Mobile Broadband System (MBS)Trends and Impacts on 60 GHz Band MMIC Devel-opment”, Electronics and Communication EngineeringJournal, pp. 187-197, June 1993.

[4] Y. Takimoto et al., “Considerations on Millimeter-WaveIndoor LAN”, Topical Symposium on Millimeter Waves,pp. 111-114, 1997.

[5] European Radiocommunication Commission, “FrequencyRange 29.7 MHz to 105 GHz and Associated EuropeanTable of Frequency Allocations and Utilizations”, ECRrep. 25, Feb. 1998; www.ero.dk.

[6] Federal Communications Commission, “Amendment ofpart 2, 15 and 97 of the Commission’s Rule to PermitUse of Radio Frequencies Above 40 GHz for New RadioApplications”, FCC 95-499, ET Docket No. 94-124, RM-8308, Dec. 15, 1995.

[7] H. Ogawa et al., “Millimeter-Wave Fiber Optics Systemsfor Personal Radio Communication”, IEEE Trans. Mi-crowave Theory Techn., pp. 2285-2293, 1992.

[8] P.M. Smith et al., “Advances in InP HEMT Technol-ogy for High Frequency Applications”, 13th International

126 BIBLIOGRAPHY

Conference on Indium Phosphide and Related Material,pp. 9-14, 2001.

[9] W.M.T. Kong et al., “Very high efficiency V-band powerInP HEMT MMICs ”, IEEE Electron Device Letters, pp.521-523, 2000.

[10] J.A. del Alamo et al., “Breakdown in Millimeter-Wave Power InP HEMT’s: A Comparison with GaAsPHEMT’s”, IEEE Journal of Solid-State Circuits, pp.1204-1211, 1999.

[11] U. Gliese et al. “Chromatic Dispersion in Fiber-Optic Mi-crowave and Millimeter-Wave Links”, IEEE Trans. Mi-crowave Theory Techn., pp. 1716-1724, 1996.

[12] A. Hilt et al. “New Carrier Generation Approach for fiber-Radio Systems to Overcome Chromatic Dispersion Prob-lems”, IEEE MTT-S Int. Microwave Symp. Digest, pp.1525-1528, 1998.

[13] D. Novak et al. “Fiber-Optic Networks for Millimeter-Wave Wireless Communications”, Asia Pacific MicrowaveConference, pp. 309-316, 1998.

[14] G.H. Smith et al. “Overcoming Chromatic-Dispersion Ef-fects in Fiber-Wireless System Incorporating ExternalModulators”, IEEE Trans. Microwave Theory Techn., pp.1410-1415, 1997.

[15] H. Schmuck “Comparison of Optical Millimeter-WaveSystem Concepts with Regard to Chromatic Dispersion”,Electronics Letters, pp. 1848-1849, 1995.

[16] N. Imai et al. “Millimeter Wave Personal Communi-cations System Using Fiber-Optic Links”, 1995 FourthIEEE International Conference on Universal PersonalCommunications, pp. 491-495, 1995.

[17] H. Kawamura et al. “High-Speed Data Transmission Us-ing Millimeter-Wave Fiber-Optic Link”, IEICE Transac-tions on Communication, pp. 1784-1791, 1996.

[18] L. Noel et al. “Novel Techniques for High-Capacity 60-

BIBLIOGRAPHY 127

GHz Fiber-Radio Transmission Systems”, IEEE Trans.Microwave Theory Techn., pp. 1416-1423, 1997.

[19] R.P. Braun et al. “Optical Millimeter-Wave System forBroadband Mobile Comminications, Devices and Tech-niques”, International Zurich Seminar on BroadbandCommunications, pp. 51-58, 1998.

[20] R.P. Braun et al. “Low-phase-noise millimeter-wave gen-eration at 64 GHz and data transmission using opticalsideband injection locking”, IEEE Photonics TechnologyLetters, pp. 728 -730, 1998.

[21] G.H. Smith et al. “Broadband Millimeter-WaveFiber-Radio Network Incorporating RemoteUp/Downconversion”, IEEE MTT-S Int. MicrowaveSymp. Digest, pp. 1509-1512, 1998.

[22] K. Kojucharow et al. “A Wireless LAN at 60 GHz -Novel System Design and Transmission Experiments”,IEEE MTT-S Int. Microwave Symp. Digest, pp. 1513-1516, 1998.

[23] M. Sauer et al. “RF-Front End and Optical Feeding Sys-tem for a Millimeter-Wave Broadband CommunicationsSyste at 60 GHz”, Microwave and Optoelectronics Con-ference, pp. 227-232, 1997.

[24] K.D. Kammeyer, “Nachrichtenubertragung”, B.G. Teub-ner Stuttgart,1992.

[25] ITU-T G.957 recomendations for SDH “Digital Sectionsand Digita Line Systems”.

[26] G. Keiser, “Optical Fiber Communications”, McGraw-Hill International Editions, 2000.

[27] Bundesamt fur Umwelt, Wald und Landschaft, “Verord-nung uber den Schutz vor nichtionisierender Strahlung(NISV)”, 1999.

[28] E. Boch “High Bandwidth mm-Wave Indoor WirelessLocal Area Networks”, Microwave Journal, pp. 152-158,January 1996.

128 BIBLIOGRAPHY

[29] D.K. Snodgrass et al. “60 GHz Radio System DesignTradeoffs”, Microwave Journal, pp. 50-62, July 2002.

[30] R. Dingle et al. “Electron mobility in modulation-dopedsemiconductor superlattices”, Appl. Phys. Lett., Vol. 33,pp. 665-667, 1978.

[31] H.L. Stormer et al. “Two dimentional electron gasat semiconductor-semiconductor interface”, Solid-StateComm., Vol. 29, pp. 705-709, 1979.

[32] S.L. Su et al. “Modulation-doped AlGaAs/GaAs FET’swith high transconductance and saturation velocity”,Electron. Lett, Vol. 18, pp. 794-796, 1982.

[33] T. Mimura et al. “A new field effect transistor with selec-tively doped GaAs/n-AlGaAs heterostructures”, Japan J.Appl. Phys., Vol. 19, L225-L227, 1980.

[34] D. Delagebeaudeuf et al. “Two-dimentional electron gasMESFET strucure”, Electron. Lett. Vol. 16, pp. 667-668,1980.

[35] A. Ketterson et al. “High transconductance In-GaAs/AlGaAs pseudomorphic modulation-doped field-effect transistor”, IEEE Electron Device Lett., Vol. EDL-6, pp. 628-630, 1985.

[36] R. Fischer et al. “An In0.15Ga0.85As/GaAs pseudomor-phic single quantum-well HEMT”, IEEE Electron DeviceLett., Vol. EDL-6, pp. 491-493, 1985.

[37] C.K. Peng et al. “Microwave performance of In-AlAs/InGaAs/InP MODFET’s”, IEEE Electron DeviceLett., Vol. EDL-8, pp. 24-26, 1987.

[38] U.K. Mishra et al. “High performance sub-micrometerAlInAs-GaInAs HEMT’s”, IEEE Electron Device Lett.,Vol. EDL-9, pp. 41-43, 1988.

[39] A. Fatimulla et al. “High performance InAlAs/InGaAsHEMT’s and MESFET’s”, IEEE Electron Device Lett.,Vol. EDL-9, pp. 328-330, 1988.

[40] M. Zaknoune et al. “InAlAs/InGaAs metamorphic

BIBLIOGRAPHY 129

HEMT with high current density and high breakdownvoltage”, IEEE Electron Device Lett., Vol. 19, pp. 345-347, 1988.

[41] D.C. Dumka et al. “MetamorphicIn0.52Al0.48As/In0.53Ga0.47As HEMTs on GaAs sub-strate with fT over 200 GHz”, Electron. Lett., Vol. 35,pp.1854-1856, 1999.

[42] M. Schefer, Integrated coplanar millimeter-wave circuitsusing InP HEMTs, Diss ETH Nr. 12363, Swiss federalinstitute of technology, Zurich, 1997.

[43] H.C. Duran, High performance InP-based HEMTs withdry etched gate recess, Diss ETH Nr. 12900, Swiss federalinstitute of technology, Zurich, 1998.

[44] IQE Ldt. (www.iqep.com), Cardiff, UK.

[45] S.R. Bahl et al. “Elimination of mesa-sidewall gate leak-age in InAlAs/InGaAs heterostructures by selective side-wall recessing”, IEEE Electron Device Lett., Vol. 13, pp.195-197, 1992.

[46] Avalon Photonics Ltd. (www.avalon.ch), Zurich, Switzer-land.

[47] P. Wolf, “Microwave properties of Schottky-barrier field-effect transistors”, IBM J. Res. Develop., Vol. 14, pp. 125-141, 1970.

[48] I. Lopez Fernadet et al. “Low-noise cryogenic X-band am-plifier using wet-etched hydrogen passiveted InP HEMTdevices”, IEEE Microwave and Guided Wave Lett., Vol.9, pp. 413-415, 1999.

[49] C.P. Wen, “Coplanar waveguide: A surface strip transmis-sion line suitable for non-reciprocal gyromagnetic deviceapplication”, IEEE Trans. Microwave Theory Techn., pp.1087-1090, 1969.

[50] W.H. Haydl et al. “Design data for millimeter wave copla-nar circuits”, 23rd European Microwave Conference, pp.223-228, 1993.

130 BIBLIOGRAPHY

[51] Sonnet, version 7.0b, Sonnet Software Inc.[52] R. Limacher, private communication, 23/11/2000.[53] M.H. Somerville, J.A. del Alamo, W. Hoke, “Direct cor-

relation between impact ionization and the kink effect inInAlAs/InGaAs HEMTs”, IEEE Electron Device Letters,vol. 17, pp. 473-475, Oct. 1996.

[54] A. Orzati, D. Schreurs, L. Pergola, H. Benedikter, O.Homan and W. Bachtold, “A 110 GHz Large-Signal Look-up Table Model for InP HEMTs Including Impact Ioniza-tion Effects. ” , to be published in the IEEE TransactionsMicrowave Theory and Technics.

[55] D.E. Root, S. Fan, J. Meyer, “Technology independentlarge signal non quasi-static FET models by direct con-struction from automatically characterized device data”,Proc. 21th European Microwave Conf., 1991, pp. 927-932.

[56] C.J. Wei, Y.A. Tkachenko, D. Bartle, “Table-based dy-namic FET model assembled from small-signal models”,IEEE Trans. Microwave Theory Techn.,vol. 47, pp. 700-705, June 1999.

[57] P.J. Tasker and J. Braunstein, “New MODFET small-signal circuit model required for millimeter-wave MMICdesign: extraction and validation to 120 GHz”, IEEEMTT-S Int. Microwave Symposium, 1995, pp. 611-614.

[58] J. Wood and D. Root, “Bias-dependent linear, scalablemillimeter-wave FET model”, IEEE MTT-S Int. Mi-crowave Symposium, 2000, pp. 1381-1384.

[59] M. Fernandez-Barciela, P.J. Tasker, Y. Campos-Roca, M.Demmler, H. Massler, E. Sanchez, M.C. Curras-Francos,M. Schlechtweg, “A simplified broad-band large-signalnonquasi-static table-based FET model”, IEEE Trans.Microwave Theory Techn.,vol. 48, pp. 395-405, March2000.

[60] Ph. Jansen, D. Schreurs, W. De Raedt, B. Nauwelaers, M.Van Rossum, “Consistent Small-Signal and Large-SignalExtractions Techniques for Heterojunction FET”, IEEE

BIBLIOGRAPHY 131

Trans. Microwave Theory and Techn., vol. 43, pp. 87-93,Jan. 1995.

[61] R. Reuter, M. Agethen, U. Auer, S. van Waasen, D.Peters, W. Brockerhoff, F.J. Tegude, “Investigation andmodeling of impact ionization with regard to the RF andnoise behaviour of HFET”, IEEE MTT-S Int. MicrowaveSymp. Digest, 1997, pp. 1317-1320.

[62] D. Root, “Analysis and exact solutions of relaxation-timedifferential equations describing non quasi-static large sig-nal FET models”, Proc. 24th European Microwave Con-ference, 1994, pp. 854-859.

[63] D.S. McPherson, K. Elgaid, I.G. Thayne, I.D. Robertson,S. Lucyszyn, “Ultra-Broadband Nonlinear PHEMT Mod-elling Using TOPASTM”, 8th IEEE International Sym-posium on High Performance Electron Devices for Mi-crowave and Optoelectronic Applications, 2000, pp. 85-88.

[64] C.J. Wei, Y.A. Tkachenko, J. Gering, D. Bartle, “Com-pletely Consistent “No-Charge” PHEMT Model Includ-ing DC/RF Dispersion”, IEEE MTT-S Int. MicrowaveSymp. Digest, 2002, pp. 2133-2136.

[65] D. Root, “Elements of Measurements-Based Large-signalDevice Modeling”, Proc. RAWCON Workshop on Model-ing and Simulation of Devices and Circuits for WirelessCommunication Systems, 1998, pp. 34.

[66] D. Root, “Nonlinear Charge Modeling for FET Large-Signal Simulations and Its Importance for IP3 and ACRPin Communications Circuits”, 2001.

[67] D. Schreurs, Y. Baeyens, B.K.J.C. Nauwelaers, W. DeRaedt, M. Van Hove, M. Van Rossum, “S-ParameterMeasurements Based Quasi-Static Large-Signal ColdHEMT Model For Resistive Mixer Design”, Int. J. Mi-crowave and Millimeter-Wave Computer-Aided Engineer-ing, Vol. 6, No. 4, pp. 250-258, 1996.

[68] G. Dambrine et al., “A New Method for Determining theFET Small-Signal Equivalent Circuit”, IEEE Trans. Mi-

132 BIBLIOGRAPHY

crowave Theory Techn.,vol. 36, pp. 1151-1159, 1988.

[69] J. Pla et al., “Nonlinear Model for Predicting Intermod-ulation Distortion in GaAs FET RF Switch Devices”,IEEE MTT-S Int. Microwave Symp. Digest, 1993, pp.641-644.

[70] D. Schreurs, A. Spiers, W. De Raedt, K. Van Der Zan-den, Y. Baeyens, M. Van Hove, B. Nauwelaers, M. VanRossum, “DC, LF dispersion and HF characterization ofshort-time stressed InP based LM-HEMTs”, Proceedingsof the 7th European Symposium on Reliability of ElectronDevices, Failure Physics and Analysis, 1996, pp. 1911-1914.

[71] L. Pergola, “Sviluppo di un modello per la simulazionedel comportamento a largo segnale di un HEMT su Fos-furo di Indio. Studio di un modello table-based e sua re-alizzazione in un simulatore circuitale”, MS Thesis 1999-2000, University of Cagliari, Supervisors: A. Orzati, G.Martines.

[72] T. McKay et al., “A High-Performance 2 − 18.5 GHzDistibuted Amplifier - Theory and Experiment”, IEEETransactions on Microwave Theory and Techniques,Vol 34, No. 12, 1986.

[73] S. Deibele et al., “Attenuation Compensation in Dis-tributed Amplifier Design” IEEE Transactions on Mi-crowave Theory and Techniques, Vol 37, No. 9, 1989.

[74] Y. Ayasli et al., “A Monolithic GaAs 1 − 13 GHzTraveling-Wave Amplifier”, IEEE Transactions on Mi-crowave Theory and Techniques, Vol 30, No. 7, 1986.

[75] J. B. Beyer et al., “MESFET Distributed Amplifier De-sign Guidelines”, IEEE Transactions on Microwave The-ory and Techniques, Vol 32, No. 3, 1984.

[76] C. Carta, “Tecnologie Microelettroniche in Fosfuro di In-dio per Amplificatori RF a Larga Banda” MS Thesis1998-1999, University of Cagliari, Supervisors: A. Orzati,L. Raffo.

BIBLIOGRAPHY 133

[77] S. A. Maas “Microwave Mixers”, Artech House, Boston,1993.

[78] A. Orzati et al., “ A 110 GHz Look-up Table Based InPHEMTs Large-Signal Model Including Impact IonizationEffects”, International Conference on Indium Phosphideand Related Materials, pp. 184 − 187, 2001.

[79] H. Zirath et al., “Analog MMICs for Millimeter-Wave Ap-plications Based on a Commercial 0.14 − µm pHEMTTechnology”, IEEE Transactions on Microwave Theoryand Techniques, Vol 49, No. 11, pp. 2086 − 2092, 2001.

[80] Virk R. S. et al., “A Comparison of W-Band MMIC Mix-ers Using InP HEMT Technology”, IEEE MTT-S Digest,pp. 435 − 438, 1997.

[81] T. Kashiwa et al., “A V-Band Monolithic InP HEMTResistive Mixer with Low LO-Power Requirement”, In-ternational Conference on Indium Phosphide and RelatedMaterials , pp. 369 − 372, 1997.

[82] Verweyen L. et al., “LMDS Up- and Down-ConverterMMIC”, IEEE MTT-S Digest, pp. 1685 − 1688, 2000.

[83] S. A. Maas, “Nonlinear Microwave Circuits”, ArtechHouse, Boston, 1988.

[84] E. O’Ciardha et al., “Generic Device Frequency Mul-tiplier Analysis-A Unified Approach”, IEEE Trans.Microwave Theory and Techn., Vol. 48, No. 7,pp. 1135 − 1141, 2000.

[85] H. Fudem et al., “Novel millimeter wave active MMICtriplers”, IEEE MTT-S Digest, pp. 387 − 390, 1998.

[86] D. G. Thomas and G. R. Branner, “Single-Ended HEMTRF/Microwave Frequency Doubler Design Using Reflec-tor Networks”, 38th Midwest Symposium on Circuits andSystems, Vol. 2, pp. 1014 − 1017, 1996.

[87] Y. Campos Roca et al., “An Optimized 25.5 − 76.5GHz PHEMT-based Coplanar Frequency Tripler”, IEEEMicrowave and Guided Wave Letters, Vol. 10, No. 6,

134 BIBLIOGRAPHY

pp. 242 − 244, June 2000.[88] E. Camargo, “Design of FET frequency multipliers and

harmonic oscillators”, Artech House, 1998.[89] A. Boudiaf et al., “A High-Efficiency and Low-Phase-

Noise 38-GHz pHEMT MMIC Tripler”, IEEE Trans.Microwave Theory and Techn., Vol. 48, No. 12,pp. 2546 − 2553, 2000.

[90] A. Orzati et al., “A 48 GHz monolithically integrated fre-quency tripler with InP HEMTs In Proc”, Indium Phos-phide and Related Materials IPRM, pages 447-450, Stock-holm, Sweden, May 2002.

[91] R. Kulke, T. Sporkmann, I. Wolff, M.J. Rosario, F.Fortes, “Active Coplanar Up-Converter For High GainV-Band Applications”, IEEE Microwave and Millimeter-Wave Monolithic Circuits Symposium, pp. 201-204, 1996.

[92] M. Schlechtweg, W. Reinert, A. Bangert, J. Braunstein,P.J. Tasker, R. Bosch, W.H. Haydl, W. Bronner, A. Huls-mann, K. Kohler, J. Seibel, R. Yu, M. Rodwell, “HighPerformance MMICs in Coplanar Waveguide Technol-ogy for Commercial V-Band and W-Band Applications”,IEEE MTT-S Digest, pp. 81-84, 1994.

[93] M. Schefer, U. Lott, H. Benedickter, Hp. Meier, W.Patrick, W. Bachtold, “Active, Monolithically IntegratedCoplanar V-Band Mixer”, IEEE MTT-S Digest, pp. 1043-1046, 1997.

[94] T. Saito, N. Hidaka, Y. Ohashi, T. Shimura, Y. Aoki,“HEMT-Based MMIC Single-Balanced Mixers for 60-GHz Indoor Comminication Systems”, GaAs IC Sympo-sium, pp. 57-60, 1993.

[95] K. Sasaki, J. Utsu, K. Matsugatani, K. Hoshino, T.Taguchi, Y. Ueno, “InP MMICs For V-Band FMCWRadar”, IEEE MTT-S Digest, pp. 937-940, 1997.

[96] A. Orzati et al., “A V-band Up-converting InP HEMTActive Mixer with Low LO-Power Requirements”, ac-cepted for publication on Microwave and Guided-Wave

BIBLIOGRAPHY 135

Letters.

[97] T. v. Kerssenbrock et al., “Flip-Chip Modules forMillimeter-Wave Sensor Applications”, 28th EuropeanMicrowave Conference, 1998, pp. 601-604.

[98] T. v. Kerssenbrock et al., “Novel 77 GHz Flip-Chip Sen-sor Modules for Automotive Radar Applications”, IEEEMTT-S Digest, pp. 289-292, 1999.

[99] www.flipchips.com

[100] P.A. Magill et al., “Flip Chip Overview”, Multi-ChipModule Conference, 1996, pp. 28-33.

[101] W. Heinrich et al., “Flip-chip interconnects for frequen-cies up to W band”, Electronics Letters,vol. 37, pp. 180-181, 2000.

[102] Finetech GmbH (www.finetech.de), Berlin, Germany.

[103] H. Kusamitsu et al., “The Flip-Chip Bump Interconnec-tion for Millimeter-Wave GaAs MMIC”, IEEE Trans. onElectronics Packaging Manufacturing , vol. 22, pp. 23-28,1999.

[104] K. Maruhashi et al., “RF Performance of a 77 GHz Mono-lithic CPW Amplifier with Flip-Chip Interconnections”,IEEE MTT-S Int. Microwave Symp. Digest, 1998, pp.1095-1098.

[105] M. Ito et al., “Analysis of Flip-Chip MMIC Structurefor 60- and 76-GHz-band Low Noise Amplifiers”, Asia-Pacific Microwave Conference, 1998, pp. 299-302.

[106] T. Hirose et al., “A Flip-Chip MMIC Design with Copla-nar Waveguide Transmission Line in the W-Band”, IEEETrans. Microwave Theory Techn., vol. 46, pp. 2276-2282,1998.

[107] R.W. Jackson et al., “Modeling Millimeter-Wave ICBehavior for Flipped-Chip Mounting Schemes”, IEEETrans. Microwave Theory Techn., vol. 45, pp. 1919-1925,1997.

[108] W. Heinrich et al., “Millimeter-Wave Characteristics of

136 BIBLIOGRAPHY

Flip-Chip Interconnects for Multichip Modules”, IEEETrans. Microwave Theory Techn., vol. 46, pp. 2264-2268,1998.

[109] H.M. Hussein et al., “An Accurate Equivalent CircuitModel of Flip-Chip and Via Interconnects”, IEEE Trans.Microwave Theory Techn., vol. 44, pp. 2543-2553, 1996.

[110] A. Jentzsch et al., “Theory and Measurements of Flip-Chip Interconnects for Frequencies up to 100 GHz”, IEEETrans. Microwave Theory Techn., vol. 49, pp. 871-877,2001.

[111] A. Jentzsch et al., “Optimization of Flip-Chip Intercon-nects for Millimeter-Wave Frequencies”, IEEE MTT-SDigest, pp. 637-640, 1999.

Curriculum Vitae

I was born in 1973 in the beautiful city of Cagliari, in Sardinia(Italy).After spending five years (1988− 1992) studying Latin, Greek,and Literature at the Liceo Classico G.M. Dettori in Cagliari,I decided I wanted to see something different and, therefore,started to study electronic engineering at the Universita degliStudi di Cagliari, where I graduated in 1998 with a thesis onthermal phenomena in electronic power devices.From 1998 to 2003, I worked as a research assistant at the Mi-crowave Electronics Laboratory of the Swiss Federal Instituteof Technology in Zurich, where I received the PhD degree inelectronic engineering.My current research interests are millimeter-wave circuit designand characterization, circuit design-oriented device modeling,and packaging.

La spiaggia del Poetto, la Torre Spagnola e la Sella del Diavolo

millimeter-wave fiber-optic communication system

Documents