A 1.93TOPS/W Deep Learning Processor with a Reconfigurable Processing Element Array Based on SRAM Access Optimization

Abstract

Deep convolutional neural networks feature numerous parameters, causing data movement to usually dominate the power consumed when computing inferences. This paper proposes an on-chip buffer access optimization method and high-data-reuse architecture that can reduce the power consumed by an on-chip buffer by up to 67.8%. The chip is designed in a TSMC 40 nm process running at 200 MHz and achieves energy efficiency of 1.93 TOPS/W.