DSP Design in Wireless CommunicationLIANG LIU AND FREDRIK EDMAN, LIANG.LIU@EIT.LTH.SE

The Evolving Wireless Scene

Range

Data

Rat

e

1m 10m 100m 1km 10km

1Kb

10Kb

100Kb

1Mb

10Mb

100Mb

Cellular (WAN)

3G Cellular

2.5 G Cellular

802.11 (LAN)

802.1a

Bluetooth (PAN)

Sensor networks

More bit/sec

More bit/($·nJ)

Courtesy: Prof. Jan Rabaey, BWRC

Evolution of High-Speed Wireless

1995 2000 2005 2010 2015

10-2

100

102

104

Mb

it/s

2Mb/s802.11

11Mb/s802.11b

54Mb/s802.11ag

600Mb/s802.11n

3.39Gb/s802.11ac 7Gb/s

802.11ad

9.6Kb/sGSM

72Kb/sGPRS

474Kb/sEDGE

2Mb/sHSDPA

84Mb/sHSPA+

150Mb/sLTE

1Gb/sLTE-A

WLANCellular

5G?

CDMA OFDM MIMO MassiveMIMO

BandwidthModulation

MIMO

Same in other communication systems

Algorithms beats Moore beats Chemists

1

10

100

1000

10000

100000

1000000

10000000Algorithmic Complexity

Processor Performance (~Moore’s Law)

Battery Capacity1G

2G

3G

Courtesy: Ravi Subramanian (Morphics)

Design Considerations

Flexible

High speed

Reliable

DESIGN TRADE-OFFS!

Small area

Low power

Low price

Time to market

Digital Hardware Platforms

Efficiency

Fle

xibi

lity

ASIC

GPPDSP

GPU

FPGA

ASIP

Optimizing DSP Implementation

Functionality

HardwarePlatform

DesignSpecification

?

Optimization

Algorithm

Architecture

Module

Circuit

CDMA

CDMA: Code Division Multiple Access

“A UMTS Baseband Receiver Chip for Infrastructure Applications”, Texas Instruments, S. Sriram et al.

Accumulator

UMTS filter in receiver systeman

t

RF ADC

Inband Outband

Reconfig. UMTS filter

De-scramble&

despreaddemod

data

desired signal“Architectural Optimization for Low Power in a Reconfigurable UMTS Filter”, in Proceedings of Wireless Personal Multimedia Communications Symposium (WPMC), Deepak Dasalukunte et al. San Diego, USA

33dB

Adaptive UMTS Filter• minimum 33dB stop band attenuation (3GPP

specification)

• required filter length of 65 taps

In a Bad ChannelD D D D

...but ONLY 5 is needed in a Good one!

OFDM

OFDM: Orthogonal Frequency Division Multiplexing

N-p

oin

t ID

FT

Pa

ralle

l to

se

rial

kx ,0

kx ,1

kNx ,1

CP

ks ,0

ks ,1

kNs ,1

OFDM

FFT/IFFT in OFDM Systems

Large number of subcarriers a large FFT, O(Nlog2N)OFDM:

• DVB-2/4/8k FFT• WLAN IEEE802.11a/g-64 FFT (48+4 subcarriers)• LTE – Long Term Evolution: 2k FFT

X1(k)

X2(k) WNk

X1(k)+WNkX2(k)

-1X1(k)-WN

kX2(k)

• Folding• Pipeline• Parallel

+

FFT: VLSI

Buttfly FIFO Mult

Multi-path delay commutator

50% 50% 50%

Signle-path delay feedback

50% 100% 50%

FFT: VLSI

High-throughput

Feature number

Technology 0.13-μm

Area 1.44mm2

Throughput 1GS/s

Power 39.6mW@409MS/s

“A 1-GS/s FFT/IFFT processor for UWB applications”, IEEE Journal of Solid-State Circuits. 2005, 40(8): 1726 - 1735

FFT: Multi-path delay feedback

• Symmetry of twiddle factors

– Reduce number of twiddle factors by 87.5%

– Only 8 possibilities for 64-point FFT

FFT: Twiddle Factors

1

2

3

4

5

6

7

8

• Shift-add for constant multiplier

ROM

ControlMUX

In Out

FFT: Twiddle Factors

• Shift-add for constant multiplier (64-P FFT)

FFT: Twiddle Factors

“A 1-GS/s FFT/IFFT processor for UWB applications”, IEEE Journal of Solid-State Circuits. 2005, 40(8): 1726 - 1735

MIMO

TransmitAntennas

ReceiveAntennas

SISO

The Radio Channel

MISO

Single Input Single Output

Multiple Input Single Output(Transmit diversity)

ReceiveAntennas

TransmitAntennas

MIMO

The Radio Channel

SIMO

Single Input Multiple Output(Receive diversity)

Multiple Input Multiple Output(Multiple data streams)

Multiple Antenna System, MIMO

• At least as many receivers as transmitted streams

• Spatial separation at both transmit and receive antennas

• Improve transmission throughput or reliability

SISO

A

MISO

AA

Interference

RL

MIMO!

RL

R

Interference!

L

Understanding MIMO via Audio

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

r = Hs + nS/P

H =

h11 h12 …….. h1N

h21 h22 …….. h2N

hM1 hM2 …….. hMN

. . …….. .

N

M

hij models fading gain between the jth

transmit

and ith

receive antenna

s rTransmitted vector Received vector

MIMO System Model

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

high-complexitySignal processing

r = Hs + n

sS/P

• Recover the transmitted signal s from the received signal r, which contains interference and noise.

Interference

Cancellation!

RL

MIMO Signal Processing - Receiver

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

Receiver Signal Processing

Symbol Detection

Channel Estimation

Matrix Manipulation

r = Hs + n

HH-1^

s = H-1r^ ^S/P

MIMO Signal Processing - Receiver

• Channel estimation: obtain the channel status by training signals

• Matrix manipulation: matrix inversion or decomposition depending on detection algorithm

• Symbol detection: estimate transmitted signal s given channel matrix H and received signal r

• Linear Detection: zero-forcing detection

• Maximum Likelihood (ML) Detection: exhaustive search

• 1.6×107 points per vector detection for 64-QAM, 4×4 MIMO

𝒓=𝑯𝒔+𝒏

��𝒛𝒇=𝑯−𝟏𝒓=𝒔+𝑯−𝟏𝒏

��𝒎𝒍=𝒂𝒓𝒈 max𝒔∈¿𝑸∨¿𝑵𝒑 (𝒔∨𝒓 ,𝑯)¿

¿𝒂𝒓𝒈 min𝒔∈¿𝑸∨¿𝑵 ¿𝒓 −𝑯𝒔∨¿𝟐 ¿¿

MIMO Signal Detection

WLAN 802.11n Example

Modulation 256QAM; 4 Tx antennas; 108 sub-channels, 4ms per OFDM symbol

ML detection 1.159 x 1017 points/sec

Current DSP technology is 1G inst/sec 108 processors!

OR (“Moores Law” ..... processor capability doubles every 18 months)

MUST WAIT 40years!

Mike Faulkner 2005, Victoria Univ.

Today…

Intel i7 CPU: 1011 inst/sec

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

Receiver Signal Processing

ML Detection

Channel Estimation

Matrix Manipulation

H

s S/P

High Complexity ML Detection

ML Detection Sphere Detection

Simplified 2D-case

Limited search space a reduced complexity

Sphere Decoding: Algorithm level optimization

Sphere Decoding: complexity reduced

Chia-Hsiang Yang, University of California, Los Angeles, 2007

Detection Performance Computational Complexity

• Near optimal ML performance with significantly reduced computational complexity (# search points)

5 10 15 20 25 3010

-5

10-4

10-3

10-2

10-1

100

SNR (dB)

BE

R

ML

Tree-PruningTree-Pruning+Reording

5 10 15 20 25 3010

0

101

102

103

104

105

SNR (dB)

Av

era

ge

# o

f S

ea

rch

Po

ints

ML

Tree-PruningTree-Pruning+Reording

44 array16-QAM

Total # search points=65536

Near ML detection with 0.1% computational complexity

• QR-Decomposition:• R upper triangular matrix

𝑯=𝑸𝑹

-1 1

-1 1 -1 1

-1 1 -1 1 -1 1 -1 1

-1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1

Pi

inci

Root

4th layer

3rd layer

2nd layer

leaf

11 12 13 141 1

2 22 23 24 2

3 333 34

4 444

0

0 0

0 0 0

R R R Ry s

y R R R s

y sR R

y sR

rQyRsys H with||||minarg 2ML

ˆ

-1 1

-1 1 -1 1

-1 1 -1 1 -1 1 -1 1

-1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1

inci

Root

4th layer

3rd layer

2nd layer

leaf

K-Best Detection, e.g., K=2

Sphere Decoding: tree-search

Design [1] Design [2] Design [3]

Gate Counter 50K 91K 491K

Throuput 136Mbps 269Mbps 1100Mbps

Sphere Decoding: VLSI

• QR-Decomposition:

– R upper triangular matrix, Q unitary matrix Q => QHQ = I

– Complexity in the order of N3 multiplications

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

Receiver Signal Processing

ML Detection

Channel Estimation

Matrix Manipulation

H

s S/P

QRD (TCAS1-11) MIMOD (ISSCC-09)

Gate Count 111K 114K

Throughput (SC/s) 12.5M 28.125M

Energy (nJ/SC/s) 12.76 5.37

QR Decomposition

Q4Q3Q2Q1H H Q1H Q2Q1H Q3Q2Q1H

Rotate by multiplying with Q1: Move all energy of first column into first element

Repeat until H becomes upper triangular.

QRD Methods

• Given’s Rotation– Works with small 2x2 rotations matrices

– Suitable for small MIMO systems

• Gram Schmidt– Operates on whole columns

– Stability problems in correlated channels

• Householder Transform– Column-wise operation

– Higher stability in correlated channels

• Householder transformation matrix

• H is an orthogonal matrix• Given vector we chose v such that

Householder Transform

Householder Transform: Example

Q4Q3Q2Q1H H Q1H Q2Q1H Q3Q2Q1H

Col 1op

Col2 op

Col3op

Col4op

H4 ... H3 ... H2 .... H1

Input matricesCol 1

opCol2 op

Col3op

Col4op

H4 ... H3 ... H2

Input matricesCol 1

opCol2 op

Col3op

Col4op

H4 ... H3

Input matricesCol 1

opCol2 op

Col3op

Col4op

H4

Input matricesCol 1

opCol2 op

Col3op

Col4op

Input matrices

Householder Transform: VLSI

• Systolic array– Suitable for pipelined data flow for example column wise matrix

operations

Anything we can do at the transmitter?

• If the receiver is not positioned directly between the speakers the received streams will be at different levels

• Equalizing at the receiver side?• Like ZF, noise enhancement

• Balance at the transmitter side• Require accurate channel information at the transmitter!

RL

L + NL, 0.5 R + NR

RL

L + NL, 2(0.5 R + NR)

2RL

L + NL, R + NR

Pre-code

MIMO Precoder: understanding via audio

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

r = Hs + n

S/P

• Zero-forcing pre-coder

𝒓=𝑯𝒙+𝒏=𝑯𝑯−𝟏 𝒔+𝒏

Tx

Tx

Tx

Tx

Data

Rx

Rx

Rx

Rx

r = Hx + n

S/P

pre

-co

din

g

xs

Low-ComplexityReceiver

[𝒓𝟏

𝒓𝟐

𝒓𝟑]=[𝟏 𝟎 𝟎

𝟎 𝟏 𝟎𝟎 𝟎 𝟏] [𝒔𝟏𝒔𝟐𝒔𝟑]+[𝒏𝟏

𝒏𝟐

𝒏𝟑]❑

Is this the best solution?

MIMO Precoder

The corresponding processor for MIMO?

Digital Signal Processors for MIMO

• Flynn’s Taxonomy– Single instruction stream, single data stream (SISD)– Single instruction stream, multiple data streams (SIMD)– Multiple instruction streams, single data stream (MISD)– Multiple instruction streams, multiple data streams (MIMD)

VLIW+SIMD

• VLIW (very long instruction word)– Instruction-level parallelism, e.g., streaming of data

• SIMD (single instruction multiple data)– Data-level parallelism, e.g., vectors

MIMO Processor Examples: Phillips EVP Vector Processor for LTE (2007)

MIMO Processor Examples: EIT Vector Processor for LTE-A (2013)

What is Next?

Goes to ‘LARGE’ Dimension?

Goes to ‘LARGE’ Dimension

• Further Scaling Up?• Size limitation in terminals• Power consumption in portable devices

Cellular Antenna WLAN Antenna

HSPA+ 2×2 802.11n 4×4

LTE 4×4 802.11ac 8×4

LTE-A 8×8 802.11? 16×16

How many antennas can we have?

2 in phone

16 in laptop [1]

24 in 80mm×80mm×80mm [2]

[1] J. B. Andersen, J. O. Nielsen, G. F. Pedersen, and G. Bauch, Channel characteristics of an indoor multiuser environment, COST 2100 TD(07) 009, 2007/Febr/26-28.[2] C-Y. Chiu, J-B. Yan, and R. D. Murch, 24-port and 36-port antenna cubes suitable for MIMO wireless communications, IEEE Trans. on Antennas and Propagation, vol. 56, no. 4, pp. 1170-1176, April 2008.

Massive MIMO or Very-Large MIMO

BS

...

A1

AM

UE1

UE2

UEK

• We think of very-large MIMO (multi-user) system• We mean M>>K>>1• We are looking for M>100 antennas!• We serve 10-20 users concurrently

F. Rusek, D. Persson, B. K. Lau, E.G. Larsson, T.L. Marzetta, O. Edfors, and F. Tufvesson, “Scaling Up MIMO: Opportunities and Challenges with Very Large Arrays”, IEEE Signal Processing Magazine, Jan. 2013

Several orders of magnitudemore transmitted-energy efficient and spectrum efficient!

Very-Large Antenna Array

Figure from Erik G. Larsson “Massive MIMO: Fundamentals, Opportunities and Challenges”

128 antenna array in Lund

Simplified processing block diagram1st ver.: LTE-like OFDM-based TRX

System architecture

System components - SDR

USRP RIO 2953R (Universal Software Radio Peripheral

• 2 RF chains

– Center frequency from 1.2 to 6 GHz

– 40MHz BW

• Xilinx Kintex-7 FPGA

– 410K logic cell

– 5Mb RAM

• ~800 MBps bidirectional data streaming

BS hardware setup: side-view

Up to 800 MB/s/direction

2.8 GB/sbidirectional

Time and Freq.synchonizationnetwork

PX

Ie-8

389

PX

Ie-8

374

PX

Ie-8

374

PX

Ie-8

374

PX

Ie-8

374

…

PX

Ie-8

374

PX

Ie-8

389

PX

Ie-8

374

PX

Ie-8

374

PX

Ie-8

374

PX

Ie-8

374

…

PX

Ie-8

374

PX

Ie-8

389

PX

Ie-8

374

PX

Ie-8

374

PX

Ie-8

374

PX

Ie-8

374

…

PX

Ie-8

374

PX

Ie-8

389

PX

Ie-667

4T

PX

Ie-8

384

PX

Ie-8

384

PX

Ie-8

384

PX

Ie-7

976

PX

Ie-8

153

LuMaMi – the Lund University Massive MIMO testbed

• World first

• 100 coherent RF chains

• Flexible architecture based on NI platform and Xilinx FPGA

• Supports 10 simultaneous single antenna users in the same time-frequency resource block

• Real time operation in the 3.7 GHz band, 20 MHz bandwidth

• 300kg and 2.5kW...

Benefits by Scaling up M• “Simplified” signal processing

– Linear processing is powerful enough when M>>K» MMSE, Zero-Forcing and Matching Filter

104 multiplications for M=100 and K=10!𝐺𝑍𝐹 𝐺𝐻 (𝑮𝑮𝐻 )−1

Low-compleixty linear pre-coding

• Neumann series based ZF pre-coding– is diagonally dominante

– NS-approximation:

Low-compleixty linear pre-coding

Possible Hardware Platform• Design Challenge

– Massive parallel processing; High data traffic; High memory bandwidth

• Interconnection: Bottleneck for real massive parallelism

– Global interconnects can be millimeters long

– Routing becomes challenging

• 3D Integration

– Increase the number of “nearest neighbors” seen by each processor

– Power and delay reduction by short vertical interconnect

Conclusions• Digital signal processing is evolving at fast pace with new

wireless technologies and applications

• “Optimal” DSP implementation is crucial to bring new DSP algorithm into practice

• “Best” design achieves balanced trade-offs depending on application requirements

• Optimization at earlier stages and do cross-layer (or co-) optimization

• Keep tracking new technologies for both algorithm and implementation

Thanks