mwoh-samos07

7/29/2019 mwoh-samos07

1/24

11

1

The Next Generation Challenge for

Software Defined Radio

Mark Woh1, Sangwon Seo1, Hyunseok Lee1, Yuan Lin1,

Scott Mahlke1, Trevor Mudge1, Chaitali Chakrabarti2, Krisztian Flautner3

1

Advanced Computer Architecture Lab, University of Michigan2Department of Electrical Engineering, Arizona State University3ARM, Ltd.


2/24

22

2

University of Michigan -SAMOS 2007

3G Wireless

Large Coverage

Outdoor - High Mobility

Up to 14Mbps


3/24

33

3


Expected Wireless Growth

The growth of wireless will require more bandwidth


4/24

44

4


4G Wireless

What we need

Adaptive high performance transmission system

Great candidate for SDR

Large Coverage 100Mbps Coverage

Outdoor - High Mobility

Macro CellsPico Cells

Isolated HotSpots 1Gbps Coverage

Indoor Very Low Mobility


5/24

55

5


Next Generation Wireless 4G

MIMO

decoder

DEMOD

(OFDM)

Channel

Decoder

MIMOencoder

MOD(OFDM)

ChannelEncoder

.

..

..

.

IFFT

FFT STBC

VBLAST

..

.

Turbo code

LDPC code

TX

RX

Antenna

.

..

3 Major Components to 4G

Modulation/Demodulation

Multiple-Input Multiple-Out (MIMO)

Channel Decoder/Encoders


6/24

66

6


Modulation - OFDM

0 fsc Nfsc

..

-Nfsc -fsc

Properties of OFDM-High Spectral Efficiency

-Low Intersymbol Intereference

-Flat Fading Subcarriers

Can sustain high data rates with

multiple users

Can be implemented with IFFT/FFT


7/2477

7


Major Component of Modulation FFT/IFFT

Very wide data level parallelism

Requires complex operations

complex

mult

x[1]

eiw

complex

add

complex

sub

x[0]

X[1]

X[0]


8/2488

8


MIMO (Multiple Input Multiple Out)

Previously we used single antenna systems

Now we use multiple antennas to increase the channel

capacity

Diversity - High Reliability

Space Time Block Codes (STBC)

Multiplexing High Throughput

Vertical-BLAST (V-BLAST)


9/2499

9


Space Time Block Codes (STBC)

CombinerChannel

Estimation

Channel

Estimation

Tx1 Tx2Transmit Antennas

x[1]

-x[2]*

x[2]

x[1]*

h11 h22

h21 h12

Channel

Rx1 Rx2

Time

Receive Antennas

n11

n12

n21

n22

~x[1]

~x[2]

h11

h12 h22

h21

Noise

y11 = h11x[1] + h12x[2] + n11

y12 = -h11x[2]* + h12x[1]* + n12

y21 = h21x[1] + h22x[2] + n21

y22 = -h21x[2]* + h22x[1]* + n22

Received

Signal

~x[1] = h11*y11 + h12y12* + h21*y21 + h22y22*

~x[2] = h12*y11 - h11y12* + h22*y21 - h21y22*


10/241010

10


STBC

Requires complex operations

Low Data Movement

Highly parallelizable

Complex

MultiplyAccumulate

Channel

Estimation

h22 h21 h12 h11

y21 y11

y22* y12*

Receiver Antenna 1 and 2

~x[1] ~x[2]

Conjugate

+Negation


11/241111

11


Vertical-BLAST (V-BLAST)

S/P

Mod

Mod

V-Blast

Detector

Demod

Demod

M Transmitters R Receivers

1 2 3 4 1 2 3 4 1 2 3 4

5 6 7 8 5 6 7 8 5 6 7 8

9 10 11 12 9 10 11 12 9 10 11 12

13 14 15 16 13 14 15 16 13 14 15 16

Data Stream of 4 TxLinear Combination

of Data

Channel

Estimation

Nulling

Vector

1 2 3 4

Subtract

Strong Signal

Repeat

Strongest

Signal


12/241212

12


V-BLAST

Implementation Based on Square Root Method for V-BLAST

Original requires repeated pseudo-inverse calculation for finding the

strongest signal

This algorithm has reduces complexity

Complexity

Requires matrix operations on complex numbers

Many Matrix Transformations


13/241313

13


Channel Decoding

3G Technologies in 4G

Viterbi

Turbo Decoder

New to 4G

LDPC

Better performance characteristics compared to Turbo and Viterbi


14/241414

14


LDPC

0 1 0 1 1 0 0 1

1 1 1 0 0 1 0 0

0 0 1 0 0 1 1 1

1 0 0 1 1 0 1 0

H =

E0

L0 L1 L2 L3

E1 E2 E3

L4 L5 L6 L7

L Node Original Value Message from Check Nodes Decision

L0

L1

L2

L3

L4

L5

L6

L7

1

1

0

1

0

1

0

1

E1 0 E3 1 1

0

0

1

0

1

0

1

E0 0 E1 0

E1 1 E2 0

E0 0 E3 1

E0 1 E3 0

E1 0 E2 1

E2 0 E3 0

E0 1 E2 1

E0

L0

L1

L2

L3

E1 E2 E3

L4

L5

L6

L7

1 1 0 01 1 0 1

E0

L0

L1

L2

L3

E1 E2 E3

0

L4

L5

L6

L7

1 1 0 01 1 0 1

Message Sent = [ 1 0 0 1 0 1 0 1 ]

Message Recieved = [ 1 1 0 1 0 1 0 1 ]


15/241515

15


LDPC

Min-Sum Decoding Used

Regular LDPC code

Can get benefit from Wide SIMD

Can do the Bit Node and Check Node

Alignment of Check and Bit nodes is a problem


16/241616

16


SODA PE Architecture

SIMD

Reg.

FileE

X

SIMD

ALU+

Mult

SIMD

Shuffle

Net-

work

(SSN)

W

B

Scalar

ALU

W

B

E

X

Scalar

RF

Local

SIMD

Memory

Local

Scalar

Memory

S

T

V

AGU

RFE

X

W

B

AGU

ALU

1. SIMD pipeline

2. Scalar pipeline

4. AGU pipeline

V

T

S

Pred.

Regs

W

B

SIMD

to

Scalar

(VtoS)ALU

RF

DMA

Global

Memory

SODA

DSP

5. DMA

3. Local

memory

SystemInterconnect

SIMD 32 Wide, 16-bit datapath, Predicate Execution


17/241717

17


Key 4Galgorithms

100 Mbps 1 GbpsMCycle/s MCycle/s

FFT 2x360 4x360

IFFT 2x360 4x360

STBC 240 -

V-BLAST - 1900

LDPC 7700 4x18500

4G Workload on SODA

100 Mbps 4G requires 8Ghz SODA PE

1 Gbps 4G requires 20Ghz SODA PE


18/241818

18


SODA With Technology Scaling

0

500

1000

1500

2000

2500

3000

3500

4000

180nm 130nm 90nm 65nm 45nm 32nm 22nm

Frequency(Mhz)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Power(W)

ITRS Scaled Frequency Fixed Scaled Frequency Scaled Power

180nm 130nm 90nm 65nm 45nm 32nm 22nm

Vdd (V) 1.8 1.3 1.1 1.1 1 0.9 0.8


19/241919

19


We cant do any of 4G with technology scaling on one core

Would 8GHz cores even be an energy efficient solution?

What about 1Gbps?

Are we ever going to get a 20GHz core?

Cannot rely on technology scaling to give us 4G for free

4G SDR will require algorithmic and architectural innovations

SDR Challenges In 4G


20/24

2020

20


4G Algorithm-Architectural Co-design

Architectural improvements (SODA II)

Specialized functional units

CISC-like complex arithmetic operations

Specialized data movement hardware

Less strain on the memory system

Wider SIMD

How wide can we go? More PEs

What does the interconnect look like?

Algorithmic optimization through parallelization

Reduce intra-kernel communication Reduce memory accesses

Arithmetic is much cheaper than data movement


21/24

2121

21


Thanks

Questions?


22/24

2222

22


Successive Cancelling for V-BLAST

V-BLAST successive interference cancelling (SIC)

The ith ZF-nulling vectorwi is defined as the uniqueminimum-norm vector satisfying

Orthogonal to the subspace spanned by the contributions to yi due

to the symbols not yet estimated and cancelled and is given by theith row of H

~

Al ti S h


23/24

2323

23


Alamouti Scheme

Assumption: the channel remains unchanged over two consecutive symbols Rate = 1

Diversity order = 2

Simple decoding

Ad t f S ft D fi d R di


24/24

24

U i it f Mi hi SAMOS 2007

Advantages of Software Defined Radio

Multi-mode operations

Lower costs

Faster time to market

Prototyping and bug fixes

Chip volumes

Longevity of platforms

Enables future wireless communication

innovations

Cognitive radio

UWB EDGE 802.16a

802.16aBluetooth

802.11b WCDMA 802.11n

SDR

mwoh-samos07

Documents