1 Gbit/s MIMO-OFDM Transmission Experiments

Transcription

1 1 Gbit/s MIMO-OFDM Trasmissio Experimets V. Jugickel, A. Forck, T. Haustei, S.Schiffermüller, C. vo Helmolt Frauhofer Ist. for Telecommuicatios Heiich-Hertz Istitute Berli, Eisteiufer 37, Germay F. Luh, M. Pollock, C. Juchems IAF GmbH Brauschweig, Berlier Str. 52J Germay M. Lampe, S. Eichiger, W. Zirwas, E. Schulz Siemes AG, ICM N PG SP RC FR Sakt-Marti Str Müche, Germay Abstract I a joit effort, we have realized a experimetal mobile commuicatio lik with a gross data rate of 1 Gbit/s, usig real-time MIMO-OFDM sigal processig. We report o the set-up of the experimetal system ad o over-the-air trasmissio experimets usig omi-directioal ateas i a idoor multi-path fadig eviromet. Keywords-No-lie-of-sight commuicatios; Spatio-temporal codig; Real-time sigal processig I. INTRODUCTION Future mobile commuicatio services are expected to support high-quality multi-media applicatios, which require substatially icreased data rates ragig from 100 Mbit/s up to 1 Gbit/s per cell. These services are expected to become available i , most likely i exteded, but defiitely limited spectrum with a scalable badwidth up to 100 MHz [1]. I order to realize the high data rates, oe may eed both: a higher badwidth ad a spectrally efficiet ew air iterface where badwidth ad effort are scalable. Multiple-iput multiple-output (MIMO) techiques with multiple ateas at the trasmitter (Tx) ad at the receiver (Rx) are attractive to ehace the spectral efficiecy. But whe they are used i combiatio with the WCDMA air iterface i curret cellular systems, the processig effort scales with the third power of the sigal badwidth, which is particularly critical at the termial side i the dow-lik. It is well kow that the effort scales just liearly with the badwidth whe the orthogoal frequecy-divisio multiplex (OFDM) techique is combied with MIMO [2]. There are sigificat efforts world-wide to further develop the MIMO- OFDM techique (see [3] ad [4]) towards applicatios. Here we describe a dedicated real-time implemetatio of the MIMO-OFDM sigal processig o a recofigurable platform with badwidth of 100 MHz. At first, we describe the Tx, the preamble used to idetify the Tx ateas at the Rx ad the chael estimatio method. The we explai how the weight matrices for the MIMO are obtaied i real-time ad how the detector is implemeted. Wherever possible, the sigal processig uses fully pipelied operatios ad massive parallel computig, to eable a cotiuous sigal flow. The MIMO- OFDM core is itegrated i a experimetal radio lik ad trasmissio experimets at 1 Gbit/s are fially reported. II. TRANSMITTER The Tx cosists of 3 parallel OFDM chais, icludig a pipelied IFFT core for each chai, implemeted i a Virtex FPGA. Processig clock is 100 MHz except the digital filterig resultig i a 200 MHz DAC clock (AD9753) of the complex base-bad sigals. Sigals are fed ito a dual-stage up-coversio RF uit usig a IF of 900 MHz ad 5.26 GHz carrier frequecy. The IQ imbalace is uced by careful matchig the DACs to the IF modulators (AD8349). Frame structure: We use a 2 ms frame with 64 OFDM symbols as a preamble for the chael estimatio, a idle gap of 16 symbols, 2400 symbols for data ad a fial gap of 20 symbols. OFDM parameters are the same as i the IEEE a stadard (64 sub-carriers, 48 used for data, 16 samples for the cyclic prefix), but the system is 5 times over-clocked. Preamble: The preamble idetifies the sigals from the three Tx ateas at the Rx. A code-multiplex approach [5] is used for the followig reasos: Firstly, we eed to maximize the eergy of each atea durig the traiig phase to uce the estimatio error. I real-time ad at very high data rates, there are additioal costraits ofte oversee: Limited time to obtai the weight matrices for the MIMO processig whe the chael chages ad limited buffer to delay the received sigals, accordigly. I practice, it is favorable to get the chael estimates istataeously after the last traiig symbol. The preamble is defied i the frequecy domai where the chip sigal p j (k) of the j th atea o sub-carrier idex ad OFDM symbol idex k is give as j p ( k) = S C ( k) (1) j where S deotes the frequecy-domai scramblig sequece adopted from the IEEE a stadard to uce the trasmitter dyamics, ad C j (t) is the j th Hadamard sequece assiged to the correspodig atea. Note that it is importat to use the same scramblig sequece for all Tx ateas i order to reuse the same chael estimator for all sub-carriers. I Fig. 1, the traiig sigals at the first ad third atea accordig to (1) are show i the frequecy-time grid of the OFDM sigals. 1 1 Actually, we have used a differet sequece o each I ad Q brach for each Tx atea i the experimetal system, due to real-valued sigal processig /05/$ IEEE 861

2 ij ij H ˆ 1 = N G G SNR = P Tx (2) Fig. 1 Graphical represetatio of the preamble accordig to (1) (:1, blue:-1, gree:0). III. CHANNEL ESTIMATION At each Rx atea, the 5.26 GHz sigals are dowcoverted to the 900 MHz IF ad the to complex base-bad. Agai, the IQ mismatch is uced by carefully matchig the AD8347 IF mixers to the AD9430 AD coverters. After aalog filterig ad 200 MHz samplig of I ad Q sigals, digital filterig formed the sample values eterig ito the MIMO- OFDM sigal processig core at 100 MHz rate. At first, the cyclic prefix is removed ad 5 pipelied FFT uits are used i parallel, oe for each Rx atea, i a Virtex2Pro/100 FPGA. Chael estimator: Chael estimatio is performed i the frequecy domai ad realized with multiple correlatio circuits operatig i parallel i the same FPGA. As explaied i [5], correlatio is performed over multiple traiig symbols usig a dedicated memory cell for each I ad Q brach for each pair of trasmit ad receive ateas ad each sub-carrier. Due to the additioal read-write operatios from ad to the memory cell, the chael estimator is revised for operatio at 100 MHz. where G is the estimator gai ad P the umber of traiig symbols P. For trial purposes, we have also implemeted a elemetary iterpolatio algorithm exploitig the fact that the chael estimates o adjacet sub-carriers are correlated. The algorithm is described i the Appedix ad it is applied separately for each pair of Tx ad Rx ateas. As show i [6], it results i a ehaced estimator gai P N G' = (3) L Tx where N is the umber of active sub-carriers ad L the umber of multi-path compoets i the chael. Fig. 3 shows a simulatio result for N=48 ad L=16. Fig. 3 Iterpolatio of raw estimates. Fig. 2 Wrapped pipelie structure used for raw chael estimatio. The ew circuit is orgaized as a wrapped pipelie usig a adder ad a dual-port RAM as show i priciple i Fig. 2. At first, the scramblig is reversed ad the sigal is either added to or subtracted from the last itermediate result recalled from port 1 of a dual-port RAM. The result is the sto i port 2. The two port addresses belog to disjoit address spaces ad addresses are couted through by the carrier idex. The idea is ow to alterately assig the address spaces from traiig symbol to traiig symbol. We use 4 Tx Rx such circuits i parallel where Tx ad Rx are the umbers of Tx ad Rx ateas, respectively. The fial estimates are trasfer ito a secod dual-port RAM after the last traiig symbol, where they ca be read asychroously ito the DSP. The raw estimates o each sub-carrier have a estimatio error of The dashed lie correspods to the true chael, from which we have got oisy raw estimates o oly 48 out of the total umber of 64 sub-carriers ( dots). After iterpolatio, the estimatio error is sigificatly uced ad the true chael is almost perfectly recostructed (full lie), also o sub-carriers where o raw estimates are available. It is obvious from (3) that this iterpolatio could be very useful i future systems to uce the legth of the pre-amble P or the umber of sub-carriers N o which raw estimatio is performed. But the iterpolatio must be performed for each pair of Tx ad Rx ateas ad it cosumes a sigificat part of the DSP processig power i our experimetal system. So it is curretly skipped i favour of a faster adaptatio to the timevariat chael. Istead we have used a loger pre-amble to improve the raw estimates. VI. ADAPTATION TO THE TIME-VARIANT CHANNEL This fuctio is implemeted outside the FPGA usig a separate TI 6713 DSP. The matrix iversio ivolved i calculatig the weight matrices for the MMSE detector may result i a huge dyamics whe the chael matrix has uced rak. This is easier hadled usig floatig-poit arithmetic. As metioed i [5], a careful optimizatio of the DSP code is madatory, ad it is further ehaced here (see Fig. 4 for curret bechmark results). With the 3 Tx ad 5 Rx ateas, 862

3 the DSP eeds roughly 0.5 ms for the 48 used carriers, so that a limited mobility ca be supported by the experimetal system. Note that additioal time (1 ms i total) is eeded to trasfer the chael estimates from the FPGA to the DSP ad the MMSE weight matrices from the DSP to the FPGA over the exteral memory iterface. So far, we have o buffer for received sigals i the FPGA, ad ew weight matrices are applied i the ext frame. Xˆ are the recostructed I ad Q braches of the Tx ateas o that carrier. The major challege is that the MMSE weight matrices are differet for each sub-carrier. So we have orde them accordig to the sub-carrier idex. The DSP stores the weight matrices i a dedicated dual-port RAM for each multiplier. Usig the secod port, the matrix-vector pipelie reads out i parallel the weights W ij at all multipliers for the curret subcarrier idex ad switches to the ext weight matrix for the ext sub-carrier at the full 100 MHz clock at which the received sigals are processed carrier-by-carrier. So the matrixvector multiplicatio uit is effectively reused for all carriers. Fig. 4 Measu times to calculate weight matrices for 48 sub-carriers. Future systems may have larger umbers of sub-carriers ad eed to adapt the sigal processig faster tha our experimetal prototype. It is a favourable property of OFDM that the requi processig power for the adaptatio to the time-variat chael scales liearly with the umber of subcarriers. I order to realize more processig power, oe might use multiple DSPs each idividually coected to the FPGA i a star cofiguratio as sketched i Fig. 5. Each DSP is the resposible for a certai subset of carriers. Based o the processig power of oe DSP, the egieer ca trade off hardware costs agaist mobility. Fig. 6 Matrix-vector multiplicatio uit used to recostruct the spatially multiplexed data streams (2 Tx, 2 Rx cofiguratio). As already metioed above, the etire MIMO-OFDM sigal processig, except the calculatio of weight matrices, is implemeted i a sigle Virtex2-Pro/100 FPGA. Sythesis results for the 3x5 (1x1) MIMO-OFDM cofiguratios are give i Table 1. It ca be observed that oly a fractio of the FPGA is occupied, which allows a efficiet routig of the VHDL desig ad reliable operatio at 100 MHz system clock. FFT Chael Estimatio Data Recostructio Total BlockRAMs 95 (19) 60 (4) 60 (4) 215 (27) of 444 Multipliers 45 (9) - 60 (4) 105 (13) of 444 Fig. 5 A star of DSPs ca be grouped aroud the FPGA to speed-up the adaptatio to the time-variat chael. V. MMSE MIMO DETECTOR The spatially multiplexed streams are recostructed i the same FPGA used for FFTs ad chael estimatio. A liear MMSE detector is implemeted usig a dedicated pipelied matrix-vector multiplicatio uit show i a simplified 2 Tx, 2 Rx cofiguratio i Fig. 6. Iput sigals Y are the frequecydomai I ad Q sigals for each Rx atea comig from the left had side. We have used a grid of 4 Tx Rx multipliers i the FPGA, followed by a pipelied cascade of adders, to perform the complete matrix-vector multiplicatio for a give sub-carrier effectively i a sigle 10 s cycle. Output sigals Slices 9878 (2401) 1272 (48) 2904 (24) (2.473) of Tab. 1 Selected sythesis data for the MIMO-OFDM sigal processig VI. CHANNEL CODING The payload is scrambled ad the split ito 4 parallel streams idividually ecoded with a covolutioal code with a costrait legth of 7 as i the IEEE a stadard. I the experimets, either the u-coded data or a chael codig with a code rate of ½ is used. The ecoded data are fed ito a pseudo-radom iterleaver. Codig is used at the first atea oly while the other two ateas are loaded with idepedet pseudo-radom data. At the receiver, the recostructed sigal costellatio of the first Tx atea is de-mapped ad the bit stream is de-iterleaved. The stream is the de-multiplexed ad 863

4 fed ito four pipelied Viterbi decoders operatig i parallel (trace-back legth 96). Parallel operatio is requi to overcome the speed limitatios of the pipelied decoder cores implemeted i a secod Virtex2Pro/100 FPGA. I priciple, it is of advatage to ecode ad iterleave the data across all ateas. I order to demostrate feasibility, however, ecodig the first stream oly may be sufficiet sice the parallel decodig cocept is obviously scalable. The MMSE MIMO detector has kowledge about the postdetectio SINR o each sub-carrier ad o each stream. Oce per frame, this iformatio is quatized ad delive from the DSP to the decoder to eable soft-decisio decodig (1 hard ad 2 soft bits). Soft decodig has a obvious effect o the performace: Beside the higher codig gai, it better exploits the multi-path diversity. VIII. TRANSMISSION EXPERIMENTS Fially, we report o over-the-air measuremets with the experimetal system i a mobile commuicatios sceario. Measuremets are coducted i a 15 7x2.80 m³ office room. At the Tx, three ateas are used addressig differet field directios (see Fig. 8, top left). The omi-directioal ateas (360 h, 25 v, 20 dow tilt) are placed at 2 m height o a wheeled photo stad. The stad is moved like a cable car usig a DC motor ad guided by a rail o a 4 m track through the room (see Fig. 8, right). At the 5.26 GHz carrier frequecy, the traslatio over about 70 carrier wavelegths forms a well reproducible chael statistics. The Rx sector ateas (65 h, 35 v) were fixed at irregular positios i a lie parallel to the Tx track ad poited approximately towards the Tx track formig a distributed atea sceario (Fig. 8, bottom). The distace betwee the Tx track ad the Rx lie was about 4 m. Fig. 7 Top left: 3-atea trasmitter uit. Top right: 5-atea receiver uit. Bottom left: Received sigal spectrum i a 100 MHz frequecy spa. Bottom right: Received 64-QAM costellatio diagrams of the three spatially data streams, averaged over all sub-carriers (The first stream is zoomed). VII. INTEGRATION The decoded data are de-scrambled ad fed ito the data sik. For the data iterface, we have used a dedicated 64-bit PCI bus iterface card betwee FPGA ad MAC cotroller realized at Tx ad Rx with two PCs based o Real-time Liux each ruig a simplified medium access cotrol (MAC) protocol stack up to the iteret protocol (IP) layer. There, the payload is bridged to a 100 Mbit/s Etheret iterface. Measuremets showed stable operatio up to 80 Mbit/s. Two otebook computers are coupled to the data source ad sik PCs via Etheret, respectively, ad a covetioal IP etwork is operated, with the uidirectioal MIMO-OFDM radio lik i betwee. The reverse lik is replaced by a Etheret cable. I this way, MPEG-ecoded HDTV video trasmissio was demostrated over the 1 Gbit/s MIMO-OFDM lik at the 3GSM World Cogress i Caes, February Fig. 7 shows the itegrated Tx ad Rx uits (top), the received sigal spectrum ad the recostructed 64-QAM costellatio diagrams of the three spatially multiplexed data streams (where the first stream is zoomed out). The costellatio diagrams are averaged over all sub-carriers. Fig. 8 Measuremet sceario. Top left: Tx atea cofiguratio. Bottom left: Rx atea cofiguratio. Right: The Tx is moved o a lie through the room. At first, we have used the built-i chael estimator to gai iformatio about the chael i the measuremet sceario. I Fig. 9, the power delay profile is show, averaged over all ateas ad sapshots alog the track. Fig. 9 Measu power delay profile Obviously, the 100 MHz badwidth resolves multiple propagatio paths i the room. These paths are a valuable source of multi-path (or frequecy-) diversity. The diversity effect ca be observed i the cumulative distributio of the chael capacity. The distributio is obtaied from measu chael estimates as follows. The estimates are corrected for the frequecy respose of Tx ad Rx chais. The mea received power is averaged over both all carriers ad all sapshots, which gives the ormalizatio factor η i (4). Fially, the capacity is obtaied as 864

5 C broadbad = 1 N 1 SNR H log 2 det 1 + HH (4) N = 0 t η the result obtaied from the slope of the u-coded BER at low atteuatio. The coded BER curves are much steeper, which is partly due to the multi-path diversity. The arrowbad capacity follows from (4) for N=1, i.e. for a sigle sub-carrier. Results are show i Fig. 10. As expected, the distributio is shifted right whe the umbers of ateas are icreased. The slope is slightly steeper ad the curve is shifted further to the right with additioal Rx ateas. The beefit of the multipath diversity is also obvious sice the broadbad distributios (full lies) are much steeper tha i the arrow-bad case (dashed lies). This effect is well kow from Mote-Carlo simulatios [4]. A simulatio curve for 3 Tx, 5 Rx with 4 paths (dotted) suggests that the chael statistics i the measuremet sceario is fairly comparable to the Rayleigh fadig model. Fig. 11 Measu average u-coded ad coded bit error rates. CONCLUSIONS Fig. 10 Compariso of broad- ad arrow-bad capacity distributios. I Fig. 11, measu average bit error rates for BPSK, 16- QAM ad 64-QAM modulatio are show. The Tx power at all ateas is set equal (±1 db) to 0 dbm at first. The automatic gai cotrol at the Rx is oce regulated at a cetral positio o the track ad held fixed for all measuremets. The Tx is moved alog the track for each value of the atteuatio. Sice a slight variatio of the mea sigal-to-oise ratio may be icluded i the data, results are plotted versus the atteuatio at each receiver. 2 Dashed lies correspod to the idividual u-coded bit error rates (BER) of the three streams, while the full lies with full circles correspod to the average BER. The full lies with squares idicate the coded BERs. At o atteuatio, the average u-coded BER i the 64- QAM mode is about 10-2 for all streams ad the average coded BER is A careful aalysis of the performace idicated that there are substatial regios alog the track where the u-coded BER of all 3 streams is below 10-2 so that all errors could be corrected. I these regios, oe may expect that the coded lik is free of errors rate if the codig is applied o all ateas. For the 3 Tx ad 5 Rx atea cofiguratio ad with the MMSE detector, oe would expect a spatial diversity order of 3 i the Rayleigh fadig chael which is fairly comparable to 2 Due to the delayed applicatio of the weight matrices, a had-shake betwee the DSP ad the AGC is requi to esure that chages i the gai are icluded i calculated weights. This had-shake was ot implemeted at the time of the measuremets. Note that the same problem pops up whe the frame sychroizatio is implemeted, which is ogoig. We have implemeted the MIMO-OFDM techique i realtime o a recofigurable processig platform ad itegrated it i a state-of-the art experimetal radio system. We have show first over-the-air trasmissio experimets at a gross data rate of 1 Gbit/s idicatig that the system is fully operatioal. Although real-time implemetatio is challegig ad several problems eed to be solved (faster adaptatio to the chael for larger umbers of sub-carriers, implemetatio of better performig MIMO detectio schemes, lik adaptatio), we feel that MIMO-OFDM is a promisig ad future-proof cadidate for ext geeratio of mobile commuicatio systems due to the favorable property that the sigal processig effort scales liearly with the badwidth. ACKNOWLEDGEMENTS This work was supported i part by the Germa Miistry of Educatio ad Research (BMBF) ad Siemes. REFERENCES [1] 3GPP TSG RAN Future Evolutio Workshop, November 2-3, Toroto, CA, see: ftp://ftp.3gpp.org/workshop/2004_11_ran_future_evo/ [2] G. G. Raleigh, J. M. Cioffi Spatio-temporal codig for wireless commuicatios, IEEE Tras. Comm. Vol. 46, No. 3, March [3] G. L. Stuber, J. R. Barry, S. W. Mclaughli, Y. G. Li, M. A. Igram, ad T. G. Pratt, Broadbad MIMO-OFDM wireless commuicatios, i Proceedigs of the IEEE, Feb. 2004, vol. 92, pp [4] A. J. Paulraj, D. A. Gore, R. U. Nabar, ad H. Bölcskei, A overview of MIMO commuicatios - A key to Gigabit wireless, Proceedigs IEEE, Vol. 92, No. 2, pp , Feb [5] V. Jugickel, T. Haustei, A. Forck, S. Schiffermueller, H. Gaebler ad C. vo Helmolt, W. Zirwas, J. Eichiger ad E. Schulz, Real-time cocepts for MIMO-OFDM, Proceedigs CIC/IEEE Global Mobile Cogress, Shaghai, Chia, October [6] T. Haustei, S. Schiffermueller, V. Jugickel, M. Schellma, T. Michel, ad G. Wuder, Iterpolatio ad Noise Reductio i MIMO-OFDM - A Complexity Drive Perspective, Proceedigs ISSPA, Sydey, Australia, Aug. 28th - 1. Sept

6 APPENDIX Here we describe a elemetary iterpolatio scheme to improve the frequecy-domai chael estimates. It is based o the elemetary relatio L 1 l H = hlexp j2π (A.1) l = 0 N betwee frequecy- ad time-domai chael coefficiets (H ad h l, respectively) i OFDM systems. (A.1) is a set of N c equatios with L variables where L is the umber of resolved multi-paths. To solve for h l, the values H are stacked i a (1xN c ) vector H where N c is the total umber of sub-carriers. Similar to the discrete Fourier trasform, (A.1) is rewritte as a multiplicatio of a (N c xl) matrix W with a vector cotaiig the time-domai chael coefficiets stacked i the (1xL) vector h H = W h (A.2) where the elemets of the matrix W are give as W l l = exp j2π. (A.3) N I practice, some values i H are ot available, due to a spectrum mask or the presece of pilot carriers. The available estimates are hece described by a uced form of (A.1) as H = W h + N (A.4) where the idex meas that the rows correspodig to the missig estimates i H ad W are filled with zeros. Also there is a vector N i (A.4) describig the estimatio error. I a first step, the time-domai estimates are obtaied by usig the pseudo-iverse W + as h ˆ + = W H (A.5) whe W is well coditioed. This is true oly for N L (iterpolatio rule). I a secod step, (A.1) is used to iterpolate the chael coefficiets. This way we get useful results also for chael coefficiets o those sub-carriers where estimates are ot available! Note that the product W W + has dimesio (N c xn c ) but rak L, which explais the filterig effect observed i Fig. 3. I the implemetatio, it may be more efficiet to use a FFT istead of (A.1), for the secod step. 866