An Energy-Efficient Deep Neural Network Training Processor with Bit-Slice-Level Reconfigurability and Sparsity Exploitation

Abstract

This paper presents an energy-efficient deep neural network (DNN) training processor through the four key features: 1) Layer-wise Adaptive bit-Precision Scaling (LAPS) with 2) In-Out Slice Skipping (IOSS) core, 3) double-buffered Reconfigurable Accumulation Network (RAN), 4) momentum-ADAM unified OPTimizer Core (OPTC). Thanks to the bit-slice-level scalability and zero-slice skipping, it shows 5.9 x higher energy-efficiency compared with the state-of-the-art on-chip-learning processor (OCLPs).