A VLSI IMPLEMENTATION OF AN ADAPTIVE-EFFORT LOW-POWER VITERBI DECODER …gallan/pdf/adaptive_viterbi.pdf · 2003-03-10 · VITERBI DECODER FOR WIRELESS COMMUNICATIONS ... •A-RSSD

A VLSI IMPLEMENTATION OF AN ADAPTIVE-EFFORT LOW-POWER

VITERBI DECODER FOR WIRELESS COMMUNICATIONS

A thesis submitted to the Department of Electrical and Computer Engineering

in conformity with the requirements forthe degree of Master of Science

Queen's UniversityKingston, Ontario, Canada

copyright Gordon John Allan, 2001

Cellular Systems• Interference based• Increasing capacity• Push towards data-centric apps

f0 f1

f0 f1f0

f0

CDMA

FDM

• Inc. speed/function à higher MIPS• Higher MIPS à low battery life

Plethora of Standards

A non-exhaustive list of international mobile standards

Peer Bluetooth North A.+ to Tetra Europe The Terrestrial Trunked Radio Peer PHS Japan Personal Handy System

AMPS North A. Advanced Mobile Phone System 1G TACS Europe Total Access Communications System

J-TACS Japan Japanese Total Access Communications System NMT Nordic Nordic Mobile Telephone CDMA North A. Code Division Multiple Access (IS-95 Standard)

2G GSM Europe Global System for Mobile communications PDC/JDC Japan Personal Digital Cellular/Japan Digital Cellular CDMAone North A. IS-95 B Standard

2.5G GPRS Europe General Packet Radio Service - GSM w/ packets Edge Europe Enhanced Data rates for Global Evolution (384 kbps) cdma2000 North A. also known as Wideband CDMA

3G UMTS Europe Universal Mobile Telecommunications System NTT DoCoMo Japan

• Many different Wireless standards• Currently have 2.5G[eneration] networks available

Error Correcting Codes (ECC)•Wireless channel is quite harsh

•Use convolutional codes to protect ‘important’ data

•Strength of the code is mainly determined by its constraint length (K=9 in 3G below)

•Power hungry to decode – especially at high data rates

Effect of Constraint Length on Error Rate

Current Solutions• Current solutions implement the Viterbi algorithm on a DSP for the

channel decoder• K=9, 1 Mbps requires

625 MIPS - 235 mA – 117% capacity - TMS320VC5441154 Mhz - 81mA – 51% capacity – SC140

• Too power hungry as rates increase to 3G• Too MIPS intensive – can’t share DSP with other operations

Convolutional Coding/Decoding

1001100 0 0

X0

X1

1

1

1+3+2+1- 1- 2- 3

t

Logic

1

0

Y0 Y1

+1 +2

Digitized Symbols

Y0 Y1

Quantized Input: +1,+2 +1, -3 +1,-2 +3,+1 +2,-0 ...

11

01

00

11

00

11

10

10

01

10

01

00

11

01

10

00

Expexted

Branch

Symbo l s

Channel Symbols: 1,1 1,0 1,0 1,0 1,1 ...Input Data: 1 0 1 1 0 ...

000

001

010

011

100

101

110

111

time(t): 0 1 2 3 4 ...

4

0

6

4

+4

+3

+1

+0 +0

+3

+0

0

3

6

+1

+2

0

+2

+1

5

3

6+0

+3

6

9

5

+3

13

5

9

1

7

9

5

7

9

9

7

3

9

7

9

3

3

5

7

7

7

11

3

9

1

9

5

9

5

9

5

7

+3

+1

+3+1

+0

+2

Expected Symbol 0,0

Expected Symbol 1,1

Reduced Complexity Decoding

• Depth First– Fano

• Metric First– Stack

• Breadth First– T-Algorithm– M-Algorithm– Reduced State Sequence Detection (RSSD)

Reduce State Sequence Detection

• Form super-states,each contain contain many sub-states

• Effectively ‘neglect’ the p-least significant bits of the Tx shift register, treating the code ‘as if’ it were of length K’ = K-p

• Cost is reduced error performance

• Complexity/Power is reduced by 2p

Further Reduction

Error Performance of RSSDFER vs SNR for RSSD in Fading Channel

1.E-05

1.E-04

1.E-03

1.E-02

1.E-01

1.E+002 4 6 8 10 12 14

SNR

FER

(102

4 bi

t fra

mes

) 481632641282565121024

Error Performance of RSSDFER vs SNR for the AWGN Channel

1.E-05

1.E-04

1.E-03

1.E-02

1.E-01

1.E+000 1 2 3 4 5

SNR

FER

481632641282565121024

Other Factors Effecting Performance

QuantizationK4, SNR 8.0 dB, 512 bit frames

0.001

0.01

0.1

11 2 3 4 5 6

Bits of Quantization

FER

Frame Size vs FER: RSSD-6 SNR=10

0.00%

0.01%

0.02%

0.03%

0.04%

0.05%

0.06%

0.07%

0.08%

0.09%

0.10%

0 500 1000 1500 2000

Frame Size (bits)

FER

Adaptive RSSD

• Normal Viterbi Requires 2 m ACS Operations

• A-RSSD “neglects” the oldest p bits of the encoder

RSSD-8(4 state)

CRC

RSSD-7(8 state)

RSSD-0(1024 state)

OK

m bits (10) = K-1

p bits (7)

Implemented: K 11=1024 State Sequence Detector

CRCOK

.

.

• Requires 2 m-p ACS operations to decode

• Cost is reduced error performance

• If decode fails, try with more effort (reduce p) CRCOK

Unrecoverable

K’ (4)

Idealized Power ConsumptionSimulation Results

Average Normalized Power (vs 4-state RSSD decode) of Adaptive RSSD with simple falloff scheduling

1

1.5

2

2.5

3

3.5

6 8 10 12 14

SNR

Nor

mal

ized

Pow

er

VLSI Implementation

Module Hierarchy

Major Implementation Considerations

• state parallel/serial• bit parallel/serial• ACS architecture• metric reduction• traceback method• survivor memory format

• source tracking• label generation• modularity• interconnect

structure

Classic Choices RSSD Related

Butterfly Structure

010

100

101

110

LSB0 feed

LSB1 feed

Butterfly StructureRSSD

0x,0y-lsb

or (if LSB1 wins)

0x,1z-lsb

1x,0y-lsb

or (if LSB1 wins)

1x,1z-lsb

x0,y

x1,z

x is binary string of length m-p-1;y,z are the p bit estimated sub-states;y-lsb, z-lsb are the strings y and z without their lsb;

label(string) produces the n branch labels when the K-bit generators are each XORed with string;

label(0x0y)

Next StateOriginal State

(superstate, substate)

label(1x0y)

label(0x1z)

label(1x1z)

010

100

101

110

LSB0 feed

LSB1 feed

RSSD Patch

Interconnect Structure

Finished Design

3200 um X 3200 um63000 cells – 0.18um TSMC(1.6 V)

17 inputs (3.3,5V), 25 outputs (3.3 V)

Area and Cell Breakdown

Power Dissipation

Power Scaling vs Decode Strength

0.10

1.00

10.00

0 100 200 300 400 500 600

States

Pow

er (m

W)

Power in mWLinear (Ideal)

Chip Specifications• natural K=11 1024 state

• rate ½, generators (3346,2751)octal

• 3 bit soft decisions

• 8 levels of reduction (p=8 to 1) from 4 to 512 superstates

• state-parallel/bit-serial architecture

• signal gating to shut-off unused states

• pipelined path computation and traceback handling

• asynchronous dual-clocks for computation and traceback

•constant traceback depth of 64 symbols

•asynchronous IO interface with handshaking and bit counters

• variable frame sizes supported; from 1 to unlimited - bit counts roll at 2056

• reconfigurable interconnect in path computation units

• 8 separate token passing networks – 1 per level of decode strength

• independent pipelined traceback units for intermediate and final decode

Quantified Results

•Compute Requires

7+p+1 clock cycles/bit (from 16 to 8 cycles)critical path – 465 Mhz

– 58 Mbps down to 29 Mbps

• Decode Requires

5 + {2(5-p) if p<5; 1 otherwise} cycles/bit (from 6 to 37 cycles) critical path (RAM) - 166 Mhz

• Overall Data rate

7.9 Mbps (RSSD-0) to 4.5 Mbps (RSSD-8)

• Output latency of 3*64 = 192 symbol periods

• Average Power~ 1.54*RSSD-8= 1.3 mW @ 1 Mbps = 0.8 mA vs 81 mA for SC140 DSP