NCPower: Power Modelling for NVM-based Neuromorphic Chip

Abstract

Spiking neural networks (SNN) on non-volatile memory (NVM) based neuromorphic computing (NC) chips have been regarded as a promising solution in power constrained scenarios, such as Internet of Things (IoT), due to its low energy consumption. The high power efficiency of NC is due to various aspects including the non-von Neumann architecture of NC chip, low power NVM, and the event driven computation of SNN etc., and introduces a large space for low power design exploration. Therefore, a comprehensive quantitative study of the power modelling for such neuromorphic computing system is important for low power design. In this work, we propose NCPower, an energy consumption estimator for NVM-based neuromorphic chip. We systemically developed analytical models based on physical laws, and verify them by comparing the analytical results with measurement results from different neuromorphic chips. We integrated NCPower in a simulator, and analyzed the accuracy and energy consumption of both the traditional multi-spike based SNN and the new single-spike based SNN. It shows that the single-spike model has 7X energy efficiency over the multi-spike model, with similar accuracy under the CIFAR-10 dataset.