Advancing energy efficiency of spiking neural network accelerator via dynamic predictive early stopping

Abstract

Spiking Neural Network (SNN) accelerators, recognized for their potential in neuromorphic processing, have yet to fully realize high energy efficiency, especially in widespread conventional non-event-based tasks, thus limiting their broader applicability. In response, we introduce an effective, cost-efficient method named dynamic predictive early stopping. This method enhances energy efficiency by predicting and stopping nearly 70% of non-essential computations during inference while ensuring near-lossless performance. The FPGA-based accelerator prototype demonstrated a 35% energy efficiency net improvement (to 208.74 GOPS/W, leading for its class). Simultaneously, the proposed method incorporates only minimal (<0.41%) extensions to hardware. With potential for further improvements and explorations, this method could bring substantial impact by being widely adopted by SNN accelerators.