A 65nm 12.92nJ/inference Mixed-Signal Neuromorphic Processor for Image Classification

Spiking neural networks are a promising candidate for next-generation machine learning and suitable powerconstrained edge devices. In this paper, we present mixed-signal neuromorphic processor that efficienctly implements an echo state network (ESN) achieves high accuracy without costly on-chip training process. The design employs charge-domain computation circuit efficiently realizes leaky integrate fire neuron. Combined with optimizing sparse connections, the proposed algorithm-hardware co-design approach results in highly energy-efficient operation while delivering accuracy. Fabricated 65nm LP process, is measured to achieve 95.35% MNIST classification accuracy, which closely matches software model, energy efficiency of 12.92nJ per classification.