(An) explainable AI accelerator with dynamic workload allocation and heat map compression/pruning

Abstract

This paper presents EPU, the first explainable AI accelerator that achieves 367FPS heat map generation for ResNet34 and state-of-the-art hardware efficiency. It introduces a new data compression format and sparsity-aware computing core for improving system performance. It proposes a dynamic inference-explanation workload allocation with a customized on-chip network to reduce external memory access by 63.7%. It also proposes point-wise gradient pruning that reduces the size of heat maps by 7.01x.