Implementation of a SNN model on an SBC-GPU and on a workstation in order to compare their efficiency

Abstract

ABSTRACT : Researchers using Spiking Neural Networks to deploy its applications on Servers and Workstations with graphics processing units because of the restricted access the specialized neuromorphic platforms. Moreover, using such conventional systems imply high energy and acquisition costs. Recently, we have seen the popularization of computing platforms with small form factors, low energy consumption, and the ability to perform artificial intelligence. These platforms, known as single-board computers, often integrate graphics processing units and other hardware accelerators; thus, they are feasible alternatives to traditional computer systems in critical energy consumption applications. This work presents our insights into implementing a 2-layer Spiking Neural Networks inference algorithm for handwritten digit recognition. We implemented the network on a GPU MALI included on the VIM3 and a workstation GPU using the C++ language and openCL. Our experimental results show that while single-board computer inference is 6x slower compared to a workstation, it is 7x more efficient in energy consumption.