Low Power Cache Design M Bilal Paracha Hisham

Low Power Cache Design M. Bilal Paracha Hisham Chowdhury Ali Raza

Acknowlegements ¡ ¡ Ching-Long Su and Alvin M Despain from University of Southern California, ”Cache Design Trade-offs for Power and Performance Optimization: A Case Study” C. L and Alvin M. Despain “ Cache Designs for Energy and Efficiency” Zhichun Zhu Xiadong Zhang, College of William and Mary, “Access Mode predictions for low-power cache design” M. D. Powell and A. Agrawal and T. N. Vijaykumar and B. Falsafi and K. Roy, Reducing Set-Associative Cache Energy via selective Direct –Mapping and Way Prediction. ”. MICRO 2001.

Today’s talk ¡ ¡ Abstract Introduction Use of cache in microprocessors Different designs to optimize cache energy and power consumption l Design Trade-offs for Power & Performance Optimization ¡ ¡ ¡ l Set-Associative Cache Energy Reduction ¡ ¡ l Way Prediction Selective direct-mapping Access Mode Prediction (AMP) ¡ ¡ ¡ Vertical Cache Partitioning Horizontal Cache Partitioning Gray Code Addressing Advantages over Way Prediction and Phased cache Different prediction techniques Evaluation Results ¡ ¡ ¡ Cache Access Times Miss Rates Cache Energy consumption

Today’s talk…. Conclusion ¡ Acknowledgements ¡

Abstract Usage of caches in modern microprocessors. ¡ Caches designed for high performance, ignore power consumption ¡ Research activities towards low power cache design ¡

Introduction Cache uses 30 -60% processor energy in embedded systems ¡ Use of caches in high performance machines ¡ Various designs to optimize energy consumption ¡

Use of cache in microprocessors ¡ High performance products go mobile (Notebooks, PDA’s etc) ¡ Cache’s as temporary storage devices ¡ Design of components with low power consumption

Designs to optimize cache energy consumption

Vertical Cache Partitioning §Block Buffer §Block Hit/Miss §Block Size

Horizontal Cache Partitioning §Cache segments §Cache sub-banks §Reduction cache accesses §Hit time, an advantage

Gray Code Addressing • Gray code vs 2’s compliment • Minimizes bit switches • 2 s Compliment: 31 bits change • Gray Code: 16 bits change

Evaluation Results ¡ ¡ ¡ ¡ ¡ <dm, 2> A direct mapped cache with block size 2 words <dm, 4> A direct mapped cache with block size 4 words <dm, 8> A direct mapped cache with block size 8 words <2 lru, 2> A 2 -way 2 words <2 lru, 4> A 2 -way words <2 lru, 8> A 2 -way words <4 lru, 2> A 4 -way words <4 lru, 4> A 4 -way words <4 lru, 8> A 4 -way words set associative cache with block size 4 set associative cache with block size 8 set associative cache with block size 2 set associative cache with block size 4 set associative cache with block size 8

Cache Access Time o. Takes less time to access direct –mapped than set associative o. Cache access of 1 K byte for dm=4. 79 ns, for set assoc=7. 15 ns o 2 way set associative is approx 50% slower than dm cache

Energy consumption vs Cache Size

Energy Consumption

Reducing Set Associative Cache Energy Via Way Prediction and Selective Direct mapping

Cache Access Energy Reduction Techniques Energy Dissipation in Data Array is much larger than in Tag Array so Energy Optimizations in Data Array only are done. ¡ Selective Direct Mapping for DCaches ¡ Way Prediction for I-Caches ¡

Different Design Techniques a) Conventional Parallel Access

b) Sequential Access

c) Way Prediction

d) Selective Direct Mapping (DM)

Prediction Framework for Selective Direct mapping (DM)

Access Mode Prediction for Low Power Cache Design

Different Cache accessing mode ¡ Phased Cache: l l Compares tag with all the tag in a particular set, If the tag matches only then, it accesses the data Consumes energy, not efficient Access the set ↓ Access all n tags ↓ Access the data corresponding to the tag

¡ Way Prediction: l l l Access only the predicted tag and data Efficient when hit rate is high Not very efficient when there is a miss (has to access rest of the tag and data elements) Access the set ↓ Way Prediction ↓ Access the predicted data and tag sub array in the set ↓ Prediction Correct Yes ↓→No Proceed Compare the rest of the data and tag array

¡ Access Mode Prediction (AMP) § § § Prediction based approach Better to use Way Prediction when hit rate is very high When hit rate is low, it is preferable to use Phased Cache approach Predicts whether cache access will result in a hit or a miss. If it predicts a hit then Way prediction is used, other wise use Phased Cache approach Accuracy of the access mode determines the efficiency of the approach

¡ Power Consumption: l l l Perfect AMP and perfect Way Prediction has a power consumption which is the lower bound of conventional set associative cache. predicted hit in the way-prediction cache, the energy consumed is Etag +Edata, compared with n × Etag+ Edata in the phased cache miss in the way-prediction cache will consume (n + 1) ×Etag + (n + 1) × Edata, in comparison with (n +1) × Etag + Edata in the phased cache.

Different Predictors ¡ Saturating Counter: l l ¡ Similar to the saturating counter of branch prediction used in project 2 Maintains a two bit counter which increments on a cache hit and decrements on a cache miss Two-level adaptive predictor: l Adaptive two level branch prediction using global pattern-history table (GAg) ¡ ¡ ¡ l Per address two level global pattern history table (PAg) ¡ ¡ K bit history register records the result of most recent K accesses For a hit register records a 1, otherwise 0 This K bit is used to index global pattern history table which has 2^K entries, each entry is a 2 bit saturation counter Each set has its own access history register All history register index a single history pattern table Correlation predictor Gshare predictor: l XOR of global access history with current reference set provides the index for global pattern history table

Misprediction rate of different predictors

Conclusion ¡ ¡ Cache Designs can be modified to obtain maximum performance and optimal energy consumption Experiments suggest that ¡ ¡ direct-mapped caches (inst and data) consume less energy for dynamic logic Set Associative consume less energy for static logic Circuit level techniques can no longer keep power dissipation under a reasonable level. Reduction of power is done on architectural level. By producing different schemes for reducing onchip cache power consumption

Questions…? ? ?