Zebra: Memory Bandwidth Reduction for CNN Accelerators With Zero Block Regularization of Activation Maps
This addresses a critical performance issue for CNN hardware accelerators, offering a significant reduction in memory bandwidth with minimal accuracy loss.
The paper tackled the memory bandwidth bottleneck in CNN hardware accelerators by proposing Zebra, a method that reduces memory bandwidth by 70% for ResNet-18 on Tiny-Imagenet with only a 1% accuracy drop.
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.