Zebra: Memory Bandwidth Reduction for CNN Accelerators With Zero Block Regularization of Activation Maps

TL;DR

This paper introduces Zebra, a method to dynamically prune unimportant activation map blocks in CNNs, significantly reducing memory bandwidth with minimal accuracy loss and easy integration with existing pruning techniques.

Contribution

Zebra is a novel zero block regularization technique that efficiently reduces memory bandwidth in CNN accelerators by dynamically pruning activation maps.

Findings

Reduces 70% of memory bandwidth for Resnet-18 on Tiny-Imagenet.

Achieves within 1% accuracy drop while maintaining high performance.

Combines effectively with Network Slimming for additional gains.

Abstract

The large amount of memory bandwidth between local buffer and external DRAM has become the speedup bottleneck of CNN hardware accelerators, especially for activation maps. To reduce memory bandwidth, we propose to learn pruning unimportant blocks dynamically with zero block regularization of activation maps (Zebra). This strategy has low computational overhead and could easily integrate with other pruning methods for better performance. The experimental results show that the proposed method can reduce 70\% of memory bandwidth for Resnet-18 on Tiny-Imagenet within 1\% accuracy drops and 2\% accuracy gain with the combination of Network Slimming.