Field Measurements of 2x2 MIMO Communications

Transcription

1 Field Measurements of x MIMO Communications Babak Daneshrad, Prof Mike Fitz, Prof UCLA EE Dept. babak@ee.ucla.edu, fitz@ee.ucla.edu Slide

2 Overview Testbed Overview Loss Due to IQ mismatch & phase noise Measurement Results MIMO Decoder ASIC Slide

3 MIMO OFDM Testbed Overview Slide 3

4 Top Level Functional Diagram Data BW = 5 MHz fc = 5. GHz FPGA AGC Trigger Data 7 ms of real-time traffic Capture fs=5 MHz 6 MB Buffer D/A TX RF RX RF A/D 6 MB Buffer Control and Config Control Graphical User Interface (GUI) Slide 4

5 Testbed Components UCLA Phase- x MIMO Testbed Memory Buffer I/O Boards Phase Locked Loop Circuit PLX Control Board Radio Freq. Circuit Slide 5

6 TX -Step RF Up-conversion PC & I/O Board 6 6 MB866 D/A (I) D/A (Q) Cheby (5th) LPF LPF AD838 DC offset DC offset AD8346 I/Q mod. DFCH3G74HDJAA BPF 747.5MHz 75MHz BW SGA-589 Driver LO 75MHz -456 RF47 AT9- HMC8MS8G AS6M- DFCB5G5LBHAA Gain PA BPF Attenuator Switch 55MHz MHz BW LO 35MHz Slide 6

7 RX step down-conversion -456 MAX7 HMC3MS8G DFCHG45HDHAA DFCB5G5LBHAA LDBG45 DFCB5G5LBHAA ML-7-S DFCHG45HDHAA C- 5.5GHz.45GHz BPF LNA BPF BPF3 LNA BPF4 BALUN Antenna 5.5GHz 5.5GHz.45GHz.45GHz MHz BW NF=.8dB NF=dB IL=.8dB MHz BW MHz BW MHz BW IL=dB IL=dB IL=dB IL=dB LO.8GHz Hz Bessel (5th) LPF VGA Bessel (5th) LPF VGA AD6644 A/D 4 I 9 5MHz / 5MHz PC & I/O Board Hz LPF VGA LPF VGA A/D 4 Q * LO.5GHz Slide 7 V AGC D/A 8 AGC start AD978 clk

8 Packet Structure Time (Length of OFDM Block) Frequency (Sub-Channel Bandwidth) A G C Coarse Sync. Fine Sync. & Channel Estmation Load Info Data & Rep. Sync. & RLS Training (Pre FFT) Time Domain Frequency Domain Processing (Post FFT) Slide 8

9 Major Impairments: Phase Noise & IQ Mismatch Phase noise mitigation is critical to OFDM Ideal Impulse Sub-carriers Power Close-in PN Sideband PN Frequency Frequency LO Frequency Frequency IQ mismatch, gain and phase, is present in all practical RF circuits I/Q mismatch causes interference from mirror subcarriers Slide 9

10 Calibration Metrics Magnitude Number of symbols x 4 Noise Signal + Noise Calculated in the frequency domain ( Signal + Noise) Energy InputSNR = log( ) Noise Energy RF and A/D Strip cyclic prefix S/P FFT FEQ P/S QAM decoding Input SNR Implementation Loss = Input SNR-Slicer SNR Slicer SNR Slide

11 SER in Perfect Timing Mode SER for 4QAM, 6QAM and 64QAM 6QAM SER 4QAM 3 64QAM 4 5 Theoretical AWGN curves Simulation Testbed Input SNR [db] Slide

12 Testbed Calibration in Perfect Timing Mode 3 Slicer SNR v/s Input SNR Implementation Loss Slicer SNR 5 5 Input SNR - Slicer SNR db Loss Due to IQ Mismach Input SNR SNR Input SNR Comparing I/Q mismatch and Phase Noise Cancellation schemes on the testbed in perfect timing mode Reference No I/Q mismatch cancellation, no Phase Noise cancellation With I/Q mismatch cancellation, no Phase Noise cancellation With I/Q mismatch cancellation, with Phase Noise cancellation Slide

13 Implementation loss under non-perfect timing 9 8 Comparing Perfect and Non perfect Timing modes Simulation Non Perfect Timing Testbed Non Perfect Timing Simulation Perfect Timing, tracking ON Testbed Perfect Timing, tracking ON Simulation Perfect Timing, no tracking Testbed Perfect Timing, no tracking 7 6 SNR Loss due to Carrier & Sampling Freq. Loops dynamics SNR Slide 3

14 Simulated Performance of MIMO-OFDM OFDM with and w/o I/Q Mismatch Cancellation SER v/s SNR for a x MIMO channel with and without IQ Mismatch 64QAM No IQ mismatch correction 6QAM optimal IQ mismatch correction 64QAM subopt IQ mismatch correction 64QAM optimal IQ mismatch correction 64QAM reference (No IQ mismatch) IQ gain mismatch Tx - (.75,.9) Tx - (.85,.95) Rx - (.85,.) Rx - (.9,.85) SER 3 6QAM subopt IQ mismatch correction IQ phase mismatch 6 degrees at both the receivers SNR 6QAM reference (No IQ mismatch) 6QAM No IQ mismatch correction 64QAM no IQ mismatch correction 64QAM subopt IQ mismatch correction 64QAM optimal IQ mismatch correction 64QAM no IQmismatch (reference) 6QAM no IQ mismatch correction 6QAM subopt IQ mismatch correction 6QAM optimal IQ mismatch correction 6QAM no IQmismatch (reference) Slide 4

15 MMSE IQ Mismatch Canceller Slide 5

16 I/Q mismatch in MIMO-OFDM OFDM systems I/Q mismatch is caused by an imbalance on the I-rail and Q-rails. This imbalance could be gain, delay or phase. Gain mismatch occurs when the amplifiers on I-rail and Q-rail have different gains. Delay mismatch occurs when the propagation delays on the two rails are different due to trace mismatches, different D/A skews, etc. Phase mismatch occurs when the sinusoids used in the I/Q modulators and demodulators are not offset by 9 degrees. I/Q mismatch can be categorized as frequency dependent or frequency dependent. Gain and phase mismatches cause frequency independent I/Q mismatch. Delay causes frequency dependent I/Q mismatch with distortion increasing on the high frequency subcarriers. Slide 6

17 Effect of I/Q mismatch in an OFDM system I/Q mismatch causes interference from the conjugate of the data on the frequency mirror sub-carrier. N N- DC Frequency Subcarriers ~ A(k) + Y(k) = B(k) X(k) + A(k) B(k) X * (N k) Slide 7

18 I/Q delay mismatch subcarrier by subcarrier EVM Image suppression 5 IQ mismatch Min Average 5 IQ phase mismatch π/64rads =.8 degrees IQ gain mismatch ±.9 db IQ combined mismatch gain = ±.9 db delay = % of Ts phase =.8 o % of Ts % of Ts -37. db -3.3 db -4.5 db db EVM 4% of Ts -5.3 db -9.4 db 5.8 o phase -3.9 db -3.9 db 3.9dB gain db db 35 4 IQ delay mismatch % of Ts IQ delay mismatch % of Ts IQ delay mismatch 4% of Ts 5 5 subcarriers Image suppression is measured by transmitting data on half the subcarriers and measuring the image strength on the mirror frequencies Slide 8

19 Effect of I/Q mismatch on a 4-QAM 4 Constellation SISO SIMOx (a).5 MIMOx (b).5 MIMO4x4 Receive diversity helps with IQ mismatch. SIMOx shows good improvement without any I/Q mismatch cancellation algorithms (c).5.5 (d) Slide 9

20 Image Suppression on the Testbed Tx Rx Tx Rx Power in db.7db Power in db.db sub carriers sub carriers Tx Rx.69 db Tx Rx.69 db Power in db Power in db sub carriers sub carriers Slide

21 Slide I/Q mismatch for MIMO I/Q mismatch for MIMO-OFDM OFDM + = ) ( * ) ( * ) ( ) ( ) ( * ) ( * ) ( ) ( ) ( * ) ( * ) ( ) ( k N V k N V k V k V k N X k N X k X k X S R Q P S R Q P S R Q P S R Q P k N Y k N Y k Y k Y For MIMO-OFDM there is interference from the conjugate of the data on the frequency mirror subcarrier of all the datastreams.

22 I/Q mismatch and EDOF EDOF measured at % outage and 3dB SNR. Capacity at % outage, 3dB SNR Image Suppression -7.7dB -.79dB -5.6dB -db EDOF Cap. EDOF Cap. EDOF Cap. EDOF Cap. x x x x x EDOF degrades slightly due to IQ mismatch!! EDOF calculations use input SNR and channel estimates ρ R C I R H H ρ = log (det( H R M + )) = log ( ) bits / s / Hz M + ε M k k= Slide A practical definition of EDOF is the difference in capacity when ρ R is doubled. EDOF=C(ρ R )-C(ρ R ) EDOF ranges from to R

23 I/Q Mismatch Cancellation on the Testbed CDF of NLOS wireless measurements.9.8 MIMOx No IQ mismatch cancellation Probability(Slicer SNR Abscissa) MIMOx with IQ mismatch cancellation 4 db at 5-percentile SIMOx No IQ mismatch cancellation SISO No IQ mismatch cancellation SISO with IQ mismatch cancellation. SIMOx with IQ mismatch cancellation SNR Slicer SNR Slide 3

24 Phase Noise Slide 4

25 Phase Noise PSD Phase Noise PSD 8dBc/Hz at Hz 4 Phase Noise [dbc/hz] 6 8 alpha= Transmitter LO (.75 GHz) alpha= 4 Theoretical Lorenzian spectrum with 3dB BW =. Hz Frequency Offset [Hz] For modeling use /f model, not /f Slide 5

26 Phase Noise with Varying FFT Sizes Probability(Slicer SNR abscissa) Phase Noise = dbc/hz at K 4 point FFT No CPE cancellation 4 point FFT With CPE cancellation No Phase Noise 64 point FFT With CPE cancellation 64 point FFT No CPE cancellation Probability(Slicer SNR abscissa) Phase Noise = dbc/hz at K 4 Oscillators Phase Noise = dbc/hz at K Oscillator Phase Noise = dbc/hz at K Oscillator Phase Noise = dbc/hz at K 4 Oscillator No Phase Noise 64 point FFT Slicer SNR Compare FFT sizes Slicer SNR Compare osc and 4 osc Common phase error (CPE) decreases with increasing FFT Size More difficult to eliminate with CPE cancellation.5 db improvement with CPE cancellation when using 64 subcarrier SISO system CPE decreases with increasing MIMO configuartion.75 db to db with x 64-point FFT Slide 6

27 Experimental Measurements Slide 7

28 Environment : Cubicle Area τ rms = 38 ns to 5 ns 39.4 m.6 m Rx Location Tx Location Transmitter location changed in 5m increments antenna placements per location.6 m 6.8 m 6.3 m 4.7 m 5. m 3.5 m 6.5 m Wall Cubicles Slide 8

29 Range measurements in the cubicle area 3 Indoor wireless range measurements 5 Input SNR 5 -dbm dbm dbm 5 mimox simox siso LOS Distance in meters 3 5 Output SNR 5 mimox 5 simox siso Distance in meters Slide 9

30 Controlled Field Trials 44.4 m Transmitter location changed in 5m increments antenna placements per location.4 m Rx Location 8 Mbps mw TX Power Tx Location Within same room (LOS) Between rooms (No LOS) Corridor (No LOS) τ rms = 5 ns τ rms = 35 ns 7.8 m Corridor 6 Mbps mw TX Power 9.7 m 9. m Room 54-4 Room m Slide 3

31 CDF of Slicer SNR for MIMOx, SIMOx and SISO CDF of Slicer SNR for MIMOx, SIMOx, SISO in LOS paths CDF of Slicer SNR for MIMOx, SIMOx, SISO in NLOS paths MIMOx.8 SIMOx Probability(Slicer SNR abscissa) SISO SIMOx Transmitter in EE 54 6 and Receiver in EE 54 6 Probability(Slicer SNR abscissa) SISO MIMOx Transmitter in EE54 4 and Receiver in EE Slicer SNR Slicer SNR Slide 3

32 Reciprocal condition numbers Information theoretic Capacity C = ρ R H R log (det( I M + H H )) = log ( + ε k ) bits / s / Hz M M R k = ρ Probability(K abscissa) CDF of reciprocal condition numbers EE54 6 (LOS) EE54 4 (NLOS) The channel matrix H is an NxM matrix with rank R. M = num of transmit antennas N = num of receive antennas ρ R = Received SNR, ε k = singular values of H Reciprocal condition number K K = min( ε max( ε k k ) ) Slide 3

33 Capacity curves in EE54-4 Capacity CCDFs using estimated channels and slicer SNR for EE54 4 (NLOS) at 3.85dB SNR Probability(Capacity abscissa) MIMOx C avg =.36 bps/hz C % =8.76 bps/hz SIMOx C avg =8.79 bps/hz C % =7.47 bps/hz SISO C avg =7.78 bps/hz C % =5.47 bps/hz SIMOx C avg =.7 bps/hz C % =9.98 bps/hz SISO C avg =. bps/hz C % =7.5 bps/hz MIMOx C avg =8.9 bps/hz C % =5.3 bps/hz Keep received power Constant for all cases By referring back to the Tx. Theoretical channel Capacity: ρ R C = log (det( I M + H M H H ) Capacity b/s/hz Capacity measured using slicer SNR Cslicer = log ( + SNRout) Slide 33

34 Optimizing Overhead Using Capacity symbols 5 symbols 5 symbols 35 symbols 5 pilot subcarriers pilot subcarriers pilot subcarriers 3.5% 9.73% 3.89% 46.5%.4% 4.8% 9.7% training pilots 5 training pilots 5 training pilots 35 training pilots Average capacity using slicer SNR 5.7 Mbps Mbps 3.9 Mbps Mbps 5 pilots 5 training pilots 5 training pilots 5 training training 5 pilots 3.39 Mbps 8.5 Mbps 3.9 Mbps Mbps Slide 34

35 RELIC An 8x8 MIMO Detection ASIC for Wideband MIMO-OFDM OFDM System Slide 35

36 Major Challenges for RELIC Wideband MIMO with high antenna count Up to 8x8, 5MHz bandwidth Dynamic reconfiguration for Different number of antennas Different antenna configurations Different FFT sizes Highly flexible packet structure support including UCLA METEOR, IEEE 8.a/g/n Slide 36

37 Wideband MIMO up to 8x8 Algorithm research RLS algorithm offers MMSE performance, fast convergence, and automatic adaptation to various channel conditions Implementation friendly architecture Systolic array RLS algorithms and architectures Frequency domain scalability Full band mode (5MHz) up to 4x4 and half band mode up to 8x8 (.5MHz) Linear interpolation in frequency domain to reduce hardware complexity Slide 37

38 Support for Different Packet Types Innovative input tagging scheme Supports different packet structures including UCLA METEOR and IEEE8.a/g/n Real-time reconfiguration of packet structure parameters such as Length of packet Number of OFDM subcarriers Length of training and retraining sequences Slide 38

39 Design Process of RELIC Design Objective Real-tim e, 8x8, 5M Hz bandw idth, 4 subcarrier Search for Available Solutions Non-Adaptive: ZF/MMSE/BLAST Adaptive: LM S/R LS Compare Performance Compare Complexity Select RLS Compare Implementation Architectures Inverse QR /QR/Extended QR Select Inverse QR Floating Point Simulation Dynamic Range Estimation Fixed Point Simulation and Word Length Optimization RTL Implementation ASIC Slide 39

40 Simulation Results MMSE vs. MMSE-VBLAST Slide 4

41 Simulation Results MMSE vs. ZF-VBLAST Slide 4

42 Required SNR (db) for Uncoded BER= -3 (QPSK) Nt Nr ZF MMSE ZF-VBLAST MMSE-VBLAST Slide 4

43 BER BER January 5 Channel RMS Delay Spread vs. Interpolation (Nc=56, x) -5 τ rms =.ns τ rms =5ns τ rms =5ns τ rms =ns τ rms =5ns τ rms =ns L= (No Interpolation) SNR (db) L= τ rms =.ns τ rms =5ns τ rms =5ns τ rms =ns τ rms =5ns τ rms =ns 8dB 34dB SNR (db) When L=, the BER floor in N c =64 case has disappeared because the cyclic prefix length is sufficiently long (56ns for N c =56) When L=, the floor rms =5ns: SNR max =34dB, BER min rms =ns: SNR max =8dB, BER min =.7% Slide 43

44 A Fully Pipelined Inverse QR-RLS RLS Architecture for OFDM S S y S S S y S3 S3 S3 S 4 S 5 S 4 S 5 S S 4 S 5 y Array internal memory stores the QR decomposition results at all pilot subcarriers S 33 S 43 S 53 y 3 Different blocks are reconnected according to the configuration mode, i.e. 8x8/half band, or 4x4/full band S S 54 y 4 S 55 y 5 Same architecture supports all different antenna setups by dynamically connecting different y 6 blocks / γ S6 S6 S7 S7 * / * / gγ gγ g γ * / S 6 S 7 g γ * 3 / S 63 S 73 g γ * 4 / S 64 S 65 S74 S75 g γ * 5 / g γ * 6 / S 66 S76 S77 g γ * 7 / y 7 Combining array has been mapped onto a linear array by timing multiplexing e a x / / e a γ W W W W 3 W 4 W 5 W 6 W 7 xˆ Slide 44

45 Topology of Different Configuration Slide 45

46 RELIC Specifications Maximum clock frequency: 5 MHz Supported antenna setup: any valid combination of antennas (Nt Nr) up to 8x8 Dual modes Full band (5MHz): up to 4x4 with 4 subcarriers Half band (.5MHz): up to 8x8 with 5 subcarriers and expandable to full band with two RELIC chips Real-time (packet-wise reconfigurable) receive antenna selection (soft switching) Extremely flexible architecture that can be easily adapted to different OFDM packet structures Slide 46

47 RELIC Die Microphotogragh Process: TSMC.8um CMOS, 3.3V/.8V Die Size: 39.4mm (core: 9.mm ) Gate Count:.3M (SRAM: 89Kb) Packaging: 8-lead PGA Power: 36mW (@58MHz, x) Clock Freq: 5MHz (max: 58MHz) Slide 47