A 28nm 15.09nJ/inference Neuromorphic Processor with SRAM-Based Charge Domain in-Memory-Computing.

Neuromorphic computing is a promising scientific field and is suitable for edge devices with constrained resources and power. However, current neuromorphic computing platforms usually rely on current-domain computing with limited accuracy and large area and energy consumption, which limit their deployment on edge devices. This article presents a low-energy neuromorphic processor that employs charge-domain computing based on Metal-Oxide-Metal (MOM) finger capacitor and SRM neuron model based on delay units. The simulation results show that our design has the lowest energy consumption per inference of 15.09nJ and inference latency of 0.46μs compared to state-of-the-art designs while the accuracy of classification 94.81% for MNIST dataset is on par with them.