Efficiency Metrics for Design Space Exploration of Wireless Baseband Processing

MPSoC’10 Nagaragawa Convention Center Gifu City, Japan

Efficiency Metrics for Design Space Exploration

of Wireless Baseband Processing

Norbert Wehn //ems.eit.uni-kl.de

Metric - Computational Requirements

Example - Mobile Phone Trends for Channel Decoding

lte

umts hsdpa

umts lte-awimax

10000 gprs wcdma cdma2000 td-scdma td-scdma gsm dmb-t dab _dvb-t dvb-s dvb-h dvb-c uwb 802.16e 802.11b 802.11n 802.11a dvd 100 1000 c operations/bit _{0.1 GOPS}

Source: Kees van Berkel, DATE2009

bluetooth blueray 1 10 0.01 0.1 1 10 100 1000 Datarate Mbps A lgorithmi c 10 GOPS cellular decoders broadcast decoders connectivity decoders

Metric – Energy Efficiency

Example - SODA, DSP and GP Architectures

3

10

LTE TC (150Mbit/s)

Area- and Energy Efficiency Design Space

better energy efficiency better overall efficiency 1 En er gy Ef fi c ien cy : o pe ra ti on /e ne rg y ( o p/ p J ) ( ) LDPC WiMedia (~1Gbit/s) CC (500Mbit/s) Min-Sum 0,1 10 100 1000 o

Area Efficiency: GOPs/mm2

-3-Min

better area efficiency

Assessment

Metrics are computation centric i.e. operations are counted only On the other hand SoC are becoming

more and more interconnect centric e.g. NoC

more and more memory centric Assessment of these metrics

Various channel decoder implementations @ research group

All architectures based on standard synthesis flows

All designs based on 65nm technology@worst case

All data in-house available

5

Overview Channel Decoders

Block-size Payload Throughput [Mbit/s] Freq. [MHz] Area [mm2] Dynamic Power [mWatt] ASIP Conv. Codes

Binary TC Duo-binary TC N=16k 40 14(6iter) 28(6iter) 385 (P&R) 0.7 (P&R) ~100

LTE Turbo LTE turbo code N=18k 150 (6iter) 300 (P&R) 2.1 (P&R) ~300 LDPC flex R=1/4 to R=9/10 N=16k 150-300(20-10iter) 385 (P&R) 1.172 (P&R) ~389 6 LDPC fixed R=3/4 N=1.2k 480 (6iter) 435 (P&R) 0.583 (P&R) ~202 LDPC

WiMedia 1.5 R=1/2-4/5 N=1.3k 640 ₉₆₀ (R=1/2,5iter)_{(R=3/4,5iter)} 265 0.51 ~193

Algorithmic Throughput Calculations [GOPs]

information bit

normalized to ~8bit addition

Infobit-Throughput

Giga operations per second [GOPs]

100Mbit/s 300Mbit/s 1 Gbit/s 100Mbit/s 300Mbit/s 1 Gbit/s CC: states=64 ~200 ~20 ~ 60 ~200 LDPC (Min-Sum) (x3.4 appr. BP) 5 iter 75/R ~7.5/R ~22.5/R ~ 75/R 10 iter 150/R ~15/R ~ 45/R ~ 150/R 20 iter 300/R ~ 30/R ~ 90/R ~ 300/R 40 iter 600/R ~ 60/R ~ 180/R ~ 600/R 7 40 iter 600/R 60/R 180/R 600/R Turbo (Max-Log) 2 iter 280 ~ 28 ~ 84 ~ 280 4 iter 560 ~ 56 ~168 ~ 560 6 iter 840 ~ 84 ~252 ~ 840 10 ASIP TC (14Mbit/s)

Area- and Energy Efficiency

1 E n er g y E ff ici ency: o pe ra ti on /e n e rg y ( o p /pJ ) LTE TC (150Mbit/s) LDPC flexible (~100 Mbit/s) LDPC WiMedia (~1Gbit/s) CC (500Mbit/s) Min-Sum 0.1 10 100 1000

Area Efficiency: GOPs/mm2

o

-3-Min

What about Memory/Data Transfers

Current metric: energy efficiency = only operations/energy

Data transfers/ accesses substantially contribute to the power consumption Example (R=0.5)

150 Mbit/s Turbo: ~126 Gops ~40 Gaccesses

150 Mbit/s LDPC : ~90 Gops ~80 Gaccesses

Efficient data transfer is key for efficient implementation E.g. LTE TC: special interleaver structure to avoid access conflicts

E.g. DVB-S2/WiMAX LDPC: special code structure to minimize access conflicts

9

Efficiency metrics based only on operations are not appropriate

 Power includes operations and accesses!

 Architectures are favored where operations dominate compared to accesses

Turbo decoder will always be evaluated to be more efficient

Same argumentation for area efficiency

E.g. DVB S2/WiMAX LDPC: special code structure to minimize access conflicts

10

LTE TC (150Mbit/s)

System Oriented Area- and Energy Efficiency

better energy efficiency better overall efficiency 1 En er g y Ef fi ci en cy : c od ed b it /en er g y ( b it /n J) ( ) LDPC WiMedia (~1Gbit/s) CC (500Mbit/s) Min-Sum

Energy efficiency : decoded information bit per energy Area efficiency : decoded information bit/s per area

10

0,1

10 100 1000

De

c

Area Efficiency: (Mbit/s)/mm2

-3-Min

better area efficiency

100

)

ASIP TC (14Mbit/s) LTE TC (150Mbit/s)

Decoders in System Design Space

1 10 En erg y Effi ciency: e code d bit/en ergy (bit /n J ) _{LTE TC (150Mbit/s)} LDPC flexible (~100 Mbit/s) LDPC WiMedia (~1Gbit/s) CC (500Mbit/s) 11 0.1 10 100 1000 10000

Area Efficiency: (Mbit/s)/mm2

d

e

Min-Sum

-3-Min

Observations

Large change in relative and absolute positions can be observed

Relative positions between Min-Sum and λ-3-Min changed

Efficiency difference between Min-Sum and λ-3-Min decreases

LDPC WiMedia decoder has a much larger efficiency than the flexible LDPC decoder

Efficiency of LDPC WiMedia decoder is not far away from CC decoder

But what about communications performance in the comparisons ?

Overall efficiency of a baseband receiver depends on

Communications performance

Implementation performance (area- and energy efficiency)

Comparison of Different Architectures

Scenario 1: Communication driven

 Comparison of two iterative decoders (TC versus LDPC)

 Identical communications performance

Different parameters (code rate, iterations) impact on implementation efficiency Scenario 2: Implementation driven

 Comparison of iterative and non-iterative decoders (LDPC versus CC)

13

p ( )

 64-state convolutional code 960 Mbit/s (WiMedia 1.2) and WiMedia 1.5 LDPC decoder Varying communication performance impact on implementations efficiency

Scenario 1: Fixed Communication Performance

14

Turbo Code

LDPC Code

10

J

)

Scenario 1: Implementation Efficiency

0.1 1 E n er gy E ff ic ien c y : ec o d e d bit /en er gy ( b it /n

TC LTE

R=1/3, 40 iter R=1/2, 20 iter R=4/5, 10 iter 15 0.01 10 100 1000

Area Efficiency: (Mbit/s)/mm2

d

TC LTE

flexible LDPC

R 1/3, 40 iter

Scenario 2: Implementation Efficiency

100 _{4dB worse com. Perf. (1iter)}

Identical Throughput 10 E n e rg y E ff ici e n cy: c o ded bi t/ e nergy (bi t/ n J

) 4dB worse com. Perf. (1iter)

5 times higher throughput Identical comm. Perf. (2iter)

Identical throughput (e.g. clock gating)

Identical Throughput (e.g. clock gating)

Identical com. Perf. (2iter) 2.5 times higher throughput

17

1

100 1000 10000 100000

Area Efficiency: (Mbit/s)/mm2

de

c

LDPC W iMedia1.5 CC W iMedia1.5

4dB better com. Perf. (5iter) Identical throughput

Summary

Understanding trade-offs between implementation efficiency,

communications performance and flexibility requirements is mandatory f ffi i b b d i

for efficient baseband receivers

Operation based metrics for energy and area efficiency can be misleading

Memory and data transfers have to be considered in metrics for design space exploration

18 Implementation efficiency metrics have to be linked to application

performance