Many-antenna base stations are interesting systems Lin Zhong http://recg.org
2
How we got started Why many-antenna base station What we have learned What we are doing now 3
How we started Why a mobile system guy got interested in massive MIMO 4
Wireless consumes a lot of power 1800 1600 1615 1400 Power (mw) 1200 1000 800 600 HTC Wizard October 2005 704 900 725 400 200 0 5 3 221 142 88 92 80 93 142 2 180 315 9 90 32 97 25 Power profile!=energy profile 5
First insight Wi-Fi more efficient than cellular MobiSys 07 6
Why is Wi-Fi more efficient? P TX = a*d 2 D 7
Horribly wasteful 8
Directional transmission! 9
Passive directional antenna to save energy (MobiCom 10) No power overhead Fixed bean patterns 10
Beamforming to save energy (MobiCom 11) Extra transceivers Steerable beams 11
Power by multi-antenna systems (uplink) P Circuit P PA =P TX / η Baseband Signal DAC Filter Mixer Filter PA 1 Frequency Synthesizer N P Shared Baseband Signal DAC Filter Mixer Filter PA N P = P shared + N P Circuit + P TX / η 12
Circuit vs. radiation power tradeoff P=P shared + 1 P Circuit + P TX / η Fixed receiver SNR
Circuit vs. radiation power tradeoff P=P shared + 2 P Circuit + P TX / η Fixed receiver SNR
Circuit vs. radiation power tradeoff P=P shared + 3 P Circuit + P TX / η Fixed receiver SNR
Circuit vs. radiation power tradeoff P=P shared + 4 P Circuit + P TX / η Fixed receiver SNR
Circuit vs. radiation power tradeoff Optimal number of antennas for efficiency N = a P /P b P
Hardware is cheap & getting cheaper P = P shared + N P Circuit + P TX / η Transmitter Power Consumption (mw) 1200 1000 800 600 400 200 SISO 2x2 MIMO 0 2002 2004 2006 2008 2010 Year Sources: IEEE Int. Solid-State Circuits Conferences (ISSCC) and IEEE Journal of Solid-State Circuits (JSSC)
Hardware is cheap & getting cheaper P = P shared + N P Circuit + P TX / η Sources: IEEE Int. Solid-State Circuits Conferences (ISSCC) and IEEE Journal of Solid-State Circuits (JSSC)
Circuit vs. radiation power tradeoff is increasingly profitable N = a P /P b P The most energy-efficient way is to use all the antennas 20
Beyond a single link 21
What the carrier wants: Use all your antennas! 22
Guiding principles distilled Spectrum is scarce Hardware is cheap, and getting cheaper 23
You can t really fit a lot of antennas in a mobile device L 24
Got a call from Erran Li, Bell Labs Spring 2011 25
3590 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER 2010 Noncooperative Cellular Wireless with Unlimited Numbers of Base Station Antennas Thomas L. Marzetta 26
Clay Shepard went to Bell Labs Summer 2011 27
Why many-antenna base station? 28
Omni-directional base station Data 1 Poor spatial reuse; poor power efficiency; high inter-cell interference 29
Sectored base station Data 1 Better spatial reuse; better power efficiency; high inter-cell interference 30
Single-user beamforming base station Data 1 Data 3 Better spatial reuse; best power efficiency; reduced inter-cell interference 31
Multi-user MIMO base station Data 2 Data 1 Data 5 M: # of BS antennas K: # of clients (K M) Best spatial reuse; best power efficiency; reduced inter-cell interference 32
Why massive? More antennas è Higher spectral efficiency More antennas è Higher energy efficiency Marzetta s key result Simple baseband technique becomes effective T.L. Marzetta. Noncooperative cellular wireless with unlimited numbers of base station antennas. IEEE Trans. on Wireless Comm., 2010. 33
How multi-user MIMO works H M: # of BS antennas K: # of clients M K 34
Multi-user MIMO: Precoding s s! = f (s, H) (Kx1 matrix) (M x 1 matrix) H M: # of BS antennas K: # of clients M K 35
Linear Precoding s (Kx1 matrix) s! = W s (M x 1 matrix) H M: # of BS antennas K: # of clients M K 36
Linear Precoding I: Zero-forcing Beamforming Null Data 1 Null Null 37
Zero-forcing Beamforming Data 2 Null Null 38
Zero-forcing Beamforming W = c H * (H T H * ) 1 Data 2 Data 1 Data 5 39
Zero-forcing does not scale well W = c H * (H T H * ) 1 Inversion of M X M matrix O(M*K 2 ) 40
Linear precoding II: Conjugate Beamforming Data 1 41
With more antennas Data 1 42
With even more antennas Data 1 43
Conjugate Multi-user Beamforming W = c H * Data 2 Data 1 Data 5 Conjugate approaches Zeroforcing as M/Kè
Conjugate scales very well W = c H * O(K) per antenna Marzetta s key result: Conjugate approaches Zeroforcing as M/Kè 45
Many-antenna vs. small cell Capital Expenditure (CAPEX) of Cell Site Major wireless equipment only 35% Just get the site to work: >50% China Mobile White Paper: C-RAN: The Road Towards Green RAN (Oct, 2011) 46
Total Cost of Ownership (TCO) Operating & Maintenance (O&M) Operating Expenditure (OPEX) The most effective way to reduce TCO is to decrease the number of sites. China Mobile White Paper: C-RAN: The Road Towards Green RAN (Oct, 2011) 47
If you ve got a site, better use as many antennas as you can 48
After a summer at Bell Labs 10-antenna prototype in the anechoic chamber at Bell Labs 49
ArgosV1 (MobiCom 12) 50
Central Controller WARP Modules Argos Interconnects Sync Distribution Argos Hub Clock Distribution Ethernet Switch51
What we have learned 52
Good news: Linear gains as # of users increases Capacity vs. K, with M = 64 53
Linear gains as # of BS antennas increases even as total P TX scaled with 1/M Capacity vs. M, with K = 15 54
Disappointment: Conjugate not approaching Zero-forcing up to 64 antennas Capacity vs. M, with K = 15 55
Disappointment: Conjugate not approaching Zero-forcing up to 64 antennas Capacity vs. M, with K = 4 Total Capacity (bps/hz) 30 25 20 15 10 Zero forcing Conjugate Local Conj. SUBF Single Ant. 5 0 20 30 40 50 60 Base Station Antennas 56
The dirty secret of massive MIMO s s! = f (s, H) (Kx1 matrix) (M x 1 matrix) H M: # of BS antennas K: # of clients M K 57
The dirty secret of massive MIMO s s! = f (s, H) (Kx1 matrix) (M x 1 matrix) H M: # of BS antennas K: # of clients M K 58
Sounding-feedback does not scale s s! = f (s, H) (Kx1 matrix) (M x 1 matrix) M: # of BS antennas K: # of clients M K 59
One must use time-division duplex and client-sent pilot s s! = f (s, H) (Kx1 matrix) (M x 1 matrix) M: # of BS antennas K: # of clients M K 60
What happens in a single coherence period Listen to pilot Send data Calculate BF weights Receive data Time Send pilot Receive data Send data Time Within coherence time 61
Both theory and our experiments only consider Listen to pilot Send data Calculate BF weights Receive data Time Send pilot Receive data Send data Time 62
What if we factor all in? Listen to pilot Send data Calculate BF weights Receive data Time Send pilot Receive data Send data Time The base station can receive during calculation but the opportunity is limited due to downlink/uplink asymmetry 63
What if we factor all in? Listen to pilot Send data Calculate BF weights Receive data Time Client mobility Channel coherence time Number of clients Time to listen to pilot Computation hardware on base station Time to calculate BF weights 64
M = 64 K = 15 Type S L Inv. Type Sym. Super Infiniband 40 Gbps 1 µs FPGA Cluster 4x10GbE 40 Gbps 20 µs 8xIntel i7 High 2x10GbE 20 Gbps 20 µs 4xIntel i7 Mid 10GbE 10 Gbps 20 µs 2xIntel i7 F Low GbE 1Gbps 20 µs Intel i7 N Zeroforcing with various hardware configurations 65
Achieved Capacity (bps/hz) 35 30 25 20 15 10 5 0 O(K) O(MK 2 ) Zero Forcing Conjugate 2 4 6 8 10 12 14 Number of Users Fixed coherence time of 30 ms with low-end hardware. 66
What we have learned Computational resources matter significantly Simplistic Conjugate beamforming works Not in Marzetta s theoretical sense Need adaptive solutions # of clients; client mobility Precoding methods: Conjugate vs. Zero-forcing 67
What we are working on 68
Going for more antennas ArgosV2 (2013) 12 WARP V3 (48 antennas) per rack Polycarbonate, dado-style shelf Anti-static spray and thermal vent Battery-powered ArgosMobile 69
96-antenna configuration
Ongoing Work: ArgosLab Software Framework for Rapid Prototyping Out-of-the-box Functionality Time/Frequency Synchronization Calibration CSI Collection Scheduled frame-based real-time Transmission
From Argos to ArgosNet 10 GbE ArgosBS 1 (Outdoor) Inter-cell interference management Pilot contamination Client grouping & scheduling Cloud RAN 10 Server GbE NetFPGA 10 GbE ArgosBS 4 (Indoor) NetFPGA Server 10 GbE NetFPGA Server 10 GbE ArgosCloud 10 GbE 10 GbE ArgosBS 2 (Outdoor) ArgosBS 3 (Outdoor) A network of massive MU-MIMO base stations 74
In summary 75
More BS antennas + MU-MIMOè Higher efficiency & lower interference Data 2 Data 1 Data 5
More BS antennas + MU-MIMOè Higher efficiency & lower interference Data 3 Data 1 Data 6 Data 12 Data 9 Data 10
Guiding Principles Spectrum is scarce Hardware is cheap, and getting cheaper 78
Acknowledgments http://argos.rice.edu 79