Practical Performance of MU-MIMO Precoding in Many-Antenna Base Stations Clayton Shepard Narendra Anand Lin Zhong Background: Many-Antennas More antennas = more capacity Traditional approaches dont scale 2 Background: Beamforming =

Destructive Interference Constructive Interference ? = 3 Background: Channel Estimation Due tothe environment andreceiver

terminalto Path Effects Align phases(Walls) at the mobility estimation has to occur ensure constructive interference quickly and periodically + BS

= + 4 Background: Channel Estimation Multiple users have to send pilots orthogonally BS 5 Frame Structure Time Division Duplex (TDD)

Uplink and Downlink use the same channel estimates (Still Coherence Retrospective) Time Channel Estimation Uplink CE Comp Downlink Computational Pipeline Uplink Overhead

Retrospectively Apply Uplink CE 6 Downlink is Limiting Factor! Background: Multi-User Beamforming

Da t a1 8 Background: Multi-User Beamforming Data 2 9 Background: Zero-forcing Null ll u

N Da t a1 Nul l N ul l Nul l 10

Background: Zero-forcing Data 2 ll u N Nu ll Null l ul N

Nu ll 11 Background: Zero-forcing Data 2 Da t ta a D a1

3 Da ta 6 Da t a 5 a t a D

4 12 Background: Scaling Up Conjugate Da t a1 13 Background: Scaling Up Conjugate Da t

a1 14 Background: Scaling Up Conjugate Da t a1 15 Background: Scaling Up Conjugate Data 2 Da t

3 a t Da a1 Da ta 6 Data 5 a D

ta 4 16 Conjugate vs. Zero-forcing Negligible Processing O(MK2) Completely Distributed Centralized

No Latency Overhead Substantial Overhead Poor Spectral Efficiency Good Spectral Efficiency 17 Under what scenarios, if any, does conjugate precoding outperform zero-forcing?

Performance Factors Environmental Complex, and constantly changing Design Straightforward and Static 19 Performance Factors Environmental Channel Coherence Precoder Spectral Efficiency Design Number of Antennas

Hardware Capability 20 Environmental Factor: Channel Coherence Coherence Time Increases frequency of channel estimation Coherence Bandwidth Increases coherence bandwidth 21 Env. Factor: Precoder Spectral Efficiency

Real-world performance, neglecting overhead Performance Depends on: User Orthogonality Propagation Effects Noise Interference Can be modeled, but impossible to capture everything 22 23 Design Factor: Number of Antennas Number of Base Station Antennas (M)

Increases amount of computation Number of User Antennas (K) Increases channel estimation and computation 24 Design Factor: Hardware Capability Conjugate has negligible computational cost Zero-forcing requires: Bi-Directional Data Transport Large Matrix Inversions 25

Zero-forcing Hardware Factors Channel Bandwidth Inversion Latency Quantization Data Transport Switching Latency Throughput 26 Performance Model

27 Conjugate vs. Zero-forcing Without Considering Computation CE Comp Transmit 29 Spectral Efficiency vs. # of BS

antennas Spectral Efficiency (bps/Hz) K = 15 # of Base Station Antennas (M) 30 Spectral Efficiency vs. # of Users Spectral Efficiency (bps/Hz) M = 64 # of Users (K)

31 Considering Computation CE Comp Transmit 32 AchievedCapacity Capacity (bps/Hz) Achieved (bps/Hz)

80 M = 64 K = 15 Conjugate 60 40 20 0 -4 10

-3 -2 10 10 Coherence Time Coherence Time (s) -1 10

(s) Zeroforcing with various hardware 33 Performance vs. # of Users M = 64 Achieved Capacity Capacity (bps/Hz) Achieved (bps/Hz) 35

Ct = 30 ms 30 25 20 15 10 Zero-Forcing Conjugate 5 0 2

4 6 8 10 # of Users (K) Number of Users 12 14 34 Max Multiplexing Gain vs. # of

Users M = 200 Ct = 30 ms 140 ZF-Super ZF-Cluster ZF-High ZF-Mid ZF-Low Conjugate Multiplexing Gain Multiplexing Gain ( * K)( K)

120 100 80 60 X: 75 Y: 46.86 X: 89 Y: 52.82 X: 58

Y: 32.39 40 X: 36 Y: 17.27 20 X: 4 Y: 1.253 0 0 20

40 60 80 Number of Users (K) # of Users (K) 100 120 140 35

Applicability Guide Base Station Design Refine model for your implementation Enables adaptive precoding 36 Ramifications More Antennas Fasteror Higher Mobility Processing Zero-forcing 1 GHz

Adaptive Precoding Conjugate 10 GHz Conclusions Accurate model of real-world precoding performance Separates unpredictable environmental factors from deterministic design Conjugate can outperform zerforcing

Useful for guiding design and enabling adaptive precoding http://argos.rice.edu 38 Questions? http://argos.rice.edu Frame Pipelining Schemes Coherence Time All Downlink

CE Com p Downlink CE Com p Coherence Time All Uplink

Downlink CE CE Coherence Time CE

Uplink (Not to Scale) Coherence Time Optimal CE Uplink Comp Downlink Uplink

CE 40