CMOS射頻前端芯片設(shè)計

1、1602IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 8, AUGUST 2010Digitally Equalized CMOS Tran smitter Fron t-E ndWith In tegrated Power Ampli?erDae Hyu n Kwon , Stude nt Member, IEEE , Hao Li , Stude nt Member, IEEE ,Yuchun Cha ng , Member, IEEE ,Richard Tse ng , Stude nt Member, IEEE , and Yun

2、 Chiu , Se nior Member, IEEEAbstract An en ergy-ef?cie nt,3.&GH z, direct-c on versi on RF tran smitter within tegrated 23-dBm, ClassB power ampli?er(PAis fabricated in a0.13-m CMOS process for the reliable tran smissio n of 20-MHz ban dwidth OFDM sig nals. Assisted by an in tegrated feedback path,

3、a two-dime nsion al, digital look-up table (2-DLUT adapted by complex gradie nt-desce nt algorithms com-pe nsates for the I/Qmismatch and memoryless nonlinearities of the whole transmit path, including the severe amplitude and phase distortio ns of the on-chip PA. Fifth-order Butterworth tran smit ?

4、lterswith 34-MHz cutoff freque ncy suf?cie ntlyatte nu ate the aliases of the 80 MS/sdigital-to-a nalog conv erters (DACs,while retai ning high in-band ?delitywithout corrupting the predistorted signal. When a 20-MHz, 64-QAM OFDM signal with 9.6-dB peak-to-average power ratio (PAPRwas tran smitted,

5、the mea-sured average drain ef?cie ncy(DEof the PA was 12.5%at 9.6dB back-off and 17.5%at 7-dB back-off, and the corresp onding error-vector magn itude (EVMwas measured to be 29.6dB and 26.3dB with equalizati on respectively. A peak DE of 55%a nd a 25-dBm saturated output power were also measured fo

6、r the same PA in a sta nd-al one package.Index Terms Adaptive equalizer, AM-AM, AM-PM, Class-B power ampli?er,CMOS RF tran smitter, digital equalizatio n, drain ef?cie ncy,erroFvector magn itude, look-up table, memoryless non-li nearities, OFDM, peak-to-average power ratio, power-added ef?cie ncy,po

7、wer ampli?er,predistorti on.I. I NTRODUCTIONWHILE digital equalizatio n has a long history of being used in com muni catio ns, its applicati on to non li nearanalog and RF circuits has been limited until recent years, mostly due to the complexity associated with a non li near formulati on, and thus,

8、 a pote ntially high impleme ntati on cost. However, as tech no logy dow nscali ng continues to reduce the digital cost dramatically and the in tegrati on of RF circuits in sta ndard CMOS process becomes com mon place, digital equalizati on has bee n in creas in gly inv estigated as a viable approac

9、h to overcome an alog and RF circuit impairme nts in deeply scaled tech no logy no des. In the last few years, high-performa nee, low-power an alog-to-digital conv ertersMa nuscript received November 23, 2009; revised February 17, 2010; ac-cepted March 16, 2010. Curre nt version published July 23, 2

10、010. This paper was 即 proved by Guest Editor Anthony Cha n Carus one. This work was sup-ported by Marvell Semic on ductor Inc. and An alog Devices Inc.The authors are with the Departme nt of Electrical and Computer En gi-neeri ng, Un iversity of Illi nois at Urba na-Champaig n, Urba na, IL 61801-230

11、7USA (e- mail:dkw on 3illi .Color versi ons of one or more of the ?gures in this paper are available on li ne at .(ADCscorrected with digital adaptive tech niq ues were widely reported 1-6 .ln the RF sector, various digital algorithms for compensating the nonlinearities of RF power ampli?ers

12、(PAshave bee n in troduced and showcaslei n either sta nd-al one or in tegrated platforms 7 -12.Drive n by high data rate and ef?cie ntuse of spectrum resource, moder ncom muni cati on sta ndards, e.g., WLAN, WiMAX, and LTE, usually feature wideba nd sig nals with a high peak-to-average power ratio

13、(PAPR.High PAPR not on ly stresses the tran smitter (TXli nearity requireme nt, thus con strai ning desig ners to use lin ear and ofte n in ef?cie ntpower ampli?ers,but also degrades the average tran smissi on ef?cie ncydue to back-off. Digital equalizati on allows the use of non li near, power-ef?c

14、ie ntPAs such as Class-B types that exhibit superior peak as well as average ef?cie ncycharacteristics compared to Class-A or -AB PAs that are curre ntly dominant in real-world applicati ons for linear ampli?cationof high PAPR signals. Fur-thermore, signal preconditioning in the digital domain makes

15、 it possible to treat the whole transmit signal path including the baseba nd buildi ng blocks, e.g., the digital-to-a nalog conv erter (DACa nd low-pass ?lter(LPF,and other RF blocks such as the mixer. Lastly, a digital treatment can be easily made adaptive, wherein wideba nd sig nals can be compe n

16、sated with less stability concern, in con trast to alter native lin earizatio n schemes such as the Cartesia n feedback 13.In practical applicati ons, non-static operat ing con diti ons due to process, temperature, and supply voltage variations make it dif?cultto model the nonlinearities of RF tran

17、smitters accu-rately, especially in a mobile en vir onment. The non li near characteristics of a PA ofte n vary widely accordi ng to its average output power, as the heat dissipation will alter the chip tem-perature and hence the behavior of active devices. The varia-tions are more signi?cantin Clas

18、sB PAs than in Class-A PAs, owing to the dyn amic n ature of the Class-B operati on and its ensuing severe non li nearity. Therefore, an adaptive treatment of the transmit-path distortions, such as the work described here, is quite desirable from a practical standpoint as it obviates the necessity o

19、f a meticulous modeling of the PA characteristics. In addition, when the adaptation is swift, the approach also pro-vides an automatic track ing capability in time-vary ing operati ng environments, making it suitable for mobile applications.In this paper, a 3.5-GHz, direct-conversion RF transmitter

20、with an on-chip 23-dBm, Class-B PA is fabricated in a0.13-m CMOS process and tested with an adaptive baseba nd digital equalizer utilizi ng a two-dime nsio nal look-up table (2-DLUT. I/Qmismatch and memoryless nonlinearities of the transmitter including the baseband circuits, i.e., the DAC and LPF,

21、and the0018-9200/$26.00?2010IEEEKWON et al. :DIGITALLY EQUALIZED CMOS TRANSMITTER FRONT-ENDWITH INTEGRATED POWER AMPLIFIERTXRNonlinear RF TXDigitalTX DataAdaptiveDigitalEqualizerError ；Linear Feedback Path1603Fig. 1. System diagram of RF transmitter with adaptive baseband digital equal- izati on.RF

22、blocks, i.e., the up-c onv ersi on mixer and PA, are uni formly treated in this prototype. Experime nts show sig ni ?ca ntimproveme nts of both in-ba nd and out-of-ba nd lin earity performa nee of the prototype CMOS tran smitter.The rema ining part of this paper is divided into ?vesect ions. Secti o

23、n II prese nts the system architecture and design consider-ations for the transmit path, the digital equalizer, and the feed-back path. Sections III and IV describe the desig n and imple-me ntatio n of the digital equalizer and various circuit blocks of the prototype in detail. The experime ntal chi

24、p and its measure-me nt results are the n covered in Sectio n V, followed by a brief con clusi on in Secti on VI.II. S YSTEM A RCHITECTURE AND C ONSIDERATIONS Fig. Ishows the simpli?edsystem diagram of the digitally equalized RF tran smitter. The feedback path, consisting of a downconversion mixer,

25、an antialiasing ?lter,and an ADC, loops the tran smitted RF sig nal back to the digital doma in for error extract ion and equalizer adaptati on. In steady state, the digital equalizer compe nsates the amplitude and phase distor-ti ons of the TX by predistort ing the baseba nd sig nal, forcing the er

26、ror power to a minimum. As a result, the transmit path is lin-earized if the feedback path is assumed linear, and the TX gain is set by the attenuation factor of the feedback path. Typically, the most visible effect of equalizati on in the freque ncy doma in is the suppressi on of spectral regrowth

27、in adjace nt cha nn els, which is exempli?edby the in termoduiao n products of the TX in Fig. 1that are can celled out by two sideba nds with equal amplitude and opposite phase produced by the equalizer. In this approach, the linearization performance of the equalizer is determ ined mostly by three

28、factors:the accuracy of the lear ned equalizer coef?cients,the ideality of the feedback path, and the severity of the TX memory effects. Desig n con siderati ons related to these aspects at the architectural level are explored in the followi ng subsect ions. A. Tran smit PathThe PA topology is perha

29、ps the most determ ining factor of the overall tran smitter power ef?cie ncydue to the high output power and lessha n-perfect ef?cie ncyof a typical RF PA. As il- lustrated in Fig. 2, the average power ef?ciencyof a clasA PAis usually severely degraded due to the large baak-f to obta in a suf?cie nt

30、lyli near operati on, especially for tran smitt ing sig nals of large PAPR, which causes the PA to operate mostly in the low output power region where its ef?ciencyis also low 14.The trade-off betwee nli nearity and ef?cie ncymotivates the use of non li near PAs such as Class-B types with a much hig

31、her ef?cie ncyin the low output power regi on (owi ngto its dyn amic bias curre nt. Large amplitude and phase distorti ons in that re-gi on, stem ming from Class-B operati on, are in evitable and will be subseque ntly treated in the digitaldomain. Moreover, a suc-cessful linearization in the saturat

32、ed power region will result in less backoff, and thus, further improve the power ef?cie ncyof the tran smitter.In general, any memory effect of the transmit path is of con-cern in a memoryless compe nsati on such as the one described here. Since one major source of memory derives from the fre-que nc

33、y resp onse of the TX LPF, the ?ltertype and order n eed to be carefully selected. Thefollow ing criteria n eed to be satis?ed in this regard.1 The LPF n eeds to pass the in termodulatio n products and harm onics gen erated by the equalizer to the RF blocks in a frequency-independent manner, i.e., w

34、ith a constant gain and linear phase response, hich typically requires a higher ?ltercut-off frequency. Any in-ba nd freque ncy-se-lective resp onse would result in a corrupti on of the sig nal prec on diti oning, thus impairi ng the compe nsati on accu-racy.2 The ?lteralso n eeds to atte nu ate the

35、 DAC aliases suf?ie ntly in order to satisfy the spectrum mask.The above desig n con siderati ons boil dow n to a trade-off be-twee n the oversampli ng ratio (OSRof the DAC and the com-plexity of the LPF. For example,aOSR is typically em-ployed in a practical 802.11a tran smitter. Without additi ona

36、l oversampli ng, a sharper LPF cutoff is esse ntially required to sat-isfy 1 and 2. Fig. 3summarizes the system-level simulatio n re-sults for various LPF types and orders, show ing the EVM perfor-ma nee and the maximum spectral comp onent in the PA output spectrum of 30MHz above and below the carri

37、er freque ncy, whe n a sta ndard OFDM sig nal with 64subcarriers and 20-MHz ban dwidth is tran smitted. Measured amplitude and phase distor-ti on curves of a sta nd-al one CMOS Class-B PA are used in this simulatio n.The results reveal that a ?fthorderButterworth ?lterachieves a reas on able bala ne

38、e betwee n in-ba nd and out-of-ba nd lin earity performa nee, displayi ng a better tha n 33-dB error-vector magn itude (EVMwhile meet ing the 802.11a spectrum mask with some margin. Elliptic and Chebyshev ?lters,especially the ?fth -orderones whose results are not show n in Fig. 3, perform worse due

39、 to their large phase variati ons n ear the cutoff freque ncy (0.5-dBi n-ba nd ripple and 25-dB stop-ba nd atte nu-ati on for ellipic ?ltersa nd 0.5-dB ripple for Chebyshev ?ltersare assumed in the simulati on. Lastly, Bessel ?lterhas in suf?-cie nt roll-off in spite of a relatively more lin ear pha

40、se resp on se. When the TX distortion is severe, e.g., a Class-C or even Class-E PA is used, high-order nonlinearities will need to be included in the treatment, and a larger OSR of the DAC would be n ecessary to accommodate the widely spread baseba nd spec-trum due to predistorti on .If the OSR is

41、not suf?cientin such a case, it would be dif?culttcmeet the in-band and out- of-ba nd1604Fig. 2. Comparison between Class-A and ClasB- RF PAs:(asimpli?edPA circuit schematic, (boperating points of the power transistors, (coutput power, and (ddrain ef?cie ncy.specs simulta neously eve n with a higher

42、-order LPF. Although this effect has bee n considered in the literature as the major lim-iting factor of the advocated equalization approach, we believe that tech no logy has evolved to a point where such a trade-off is worthwhile or eve n ben e?cial.After all, data conv erters have ben e?tedfrom te

43、ch no logy scali ng much more tha n mon olithic RF tran smitters (withi ntegrated PAs have, especially those im-pleme nted in silic on tech no logy.B. TX Error Mecha ni sms and Equalizati onWhereas the memoryless amplitude and phase errors of a PA are mostly amplitude- oriented 14thusan amplitude-ad

44、-dressed, or one-dimensional LUT with complexcoef?cients卻 刑 黨 53 宛 T * 鋼C uwff F oqusnci (UMi):;11 Hultvrworir Jrd bltrwrlh Sih -T &hpb忙24 2B 3236 4Q 44F recjutn專Mtitffla-0 亠isIEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 8, AUGUST* Brfterwrtjrth W : BHAamrlh Sfhr- 鼻f J EHrtuMChtfi 艸血：M 也亍 丸 奄畝

45、 k Aci1 Ft&qww?r (W2010Fig. 3. Effects of TX ?ltertype and order on (aoutof-ba nd and (bin-ba ndlin earizati on performa nee, wherePDis de?n edas the maximum spectral comp onent in the PA output spectrum of 30MHz above and below f (802.11ais assumed.suf?cie ntfor the compe nsatio n an extnded treatm

46、e nt of the en tire tran smit path in clud ing DAC, LPF, and mixer n ecessitates an I-a nd Q-addressed, or 2-D LUT. The dominant memoryless error mecha ni sms of a typical RF tran smissi on process and the ensuing distorti ons of the baseba nd I/Qpla ne are illustrated in Fig. 4and expla ined as fol

47、lows.1 The non li near transcon ducta nee of the PA in put tran sis-tors un der large-sig nal excitati on causes the env elope of the RF output waveform to differ from the corresp onding baseba nd in put, which is known as the AM-AM effect; the sig nal- depe ndent capacita nee of the tran sistors al

48、so leads to amplitude-depe ndent phase resp on ses of the RF sig nal known as the AM-PM effect. Their impacts on the base-ba nd I/Qpla ne are rotati on ally symmetric.2 The nonlinearity of the baseband I/Qpaths, from the DAC to the mixer, distorts the baseba nd I/Qpla ne, which dis-plays both phase

49、and amplitude depe nden cies. The distortion pattern, as shown in Fig. 4, can be divided into fourKWON et al. :DIGITALLY EQUALIZED CMOS TRANSMITTER FRONT-ENDWITH INTEGRATED POWER AMPLIFIERRF AM-AM & AM-PMoriginalM disiorted一 DigitalTX DMANonlinear RF Transmittercombi n名 dLO l/Q mismatch1605Fig. 4. D

50、istortio n effects of the baseba nd l/Qpla ne due to various TX memoryless errors.ide ntical quadra nts assu ming that the I/Qpaths are well matched and eve n-o rder distortio ns are n egligible.3 The gain mismatch of the baseba nd I/Qpaths, and the am-plitude and phase mismatch betwee n the quadrat

51、ure LO sig nals scale and rotate the baseba nd I/Qpla ne, which also displays phase as well as amplitude depe nden cies. Note that the delay mismatch of the baseband I/Qpaths is a memory effect and thus is not treated in the memoryless equalizer; it can be largely avoided by a symmetric layout due t

52、o the low freque ncy n ature of the baseba nd sig nals. In stead of ide ntify ing and compe nsat ing the above three types of errors independently, they are equally (blindlytreated in a 2-D LUT in this work, for the sake of compact ness as well as adaptatio n speed. In additi on, a 2-D LUT can also

53、treat the even-order distortions stemming from the circuit offset, mismatch, and large-signal operation of the devices in various stages of the transmit path.For coef?cie ntadaptatio n, the equalizer should have the caability of ide ntify ing the TX memoryless errors by compari ng the origi nal data

54、 and the loopback samples (provisi on edby the feedback path, which are pote ntially delayed, phase rotated, and eve n corrupted by various impairme nts such as memory effect, on-chip coupli ng, DC offset, quantization and circuit noises. The equalization algorithm and implementation details regardi

55、 ng these issues will be covered in Secti on III. C. Feedback PathThe feedback loop in here nt to the equalizer is composed of the tran smitter and a dedicated receiver (RXcascaded. Ide-ally, the compe nsated TX can only be as lin ear as the RX path, which places a stringent linearity constraint on

56、the RX circuits. In a time- divisio n duplex (TDDsystem, the no rmal receiver can pote ntially be time-shared as the feedback path. However, much simpler feedback circuits can be employed whe n they are dedicated to the purpose of equalization; i.e., a high RX sensi-tivity is unnecessary as the PA o

57、utput is directly coupled intothe feedback path, where the sig nal is so strong that any un-wan ted in terfere nces coupled in from the antenna are esse ntially wiped out duri ng the equalizer training. Eve n in the case when a strong interferer is present, it still tends to be averaged out in the l

58、east- mea n-square (LMSiteratio n loop of the equalizer, because it is un correlated with the transmitted symbols. In ad-dition, the signal-tonoise ratio (SNRand noise ?gure(NFof the RX are much relaxed as well, since random noise tends to be averaged out in a similar manner. For example, the resolutions of the TX and RX quantization in an 802.11a system for 1%steady-state mean-square error (MSEat TX output are 8-bit and 6-bit, respectively, as prese nted in the compa nion paper 16.Note that the ADC resolutio n is relaxe