Evaluating the impact of optical interconnects on a multi-chip machine-learning architecture

Abstract

Following trends that emphasize neural networks for machine learning, many studies regarding computing systems have focused on accelerating deep neural networks. These studies often propose utilizing the accelerator specialized in a neural network and the cluster architecture composed of interconnected accelerator chips. We observed that inter-accelerator communication within a cluster has a significant impact on the training time of the neural network. In this paper, we show the advantages of optical interconnects for multi-chip machine-learning architecture by demonstrating performance improvements through replacing electrical interconnects with optical ones in an existing multi-chip system. We propose to use highly practical optical interconnect implementation and devise an arithmetic performance model to fairly assess the impact of optical interconnects on a machine-learning accelerator platform. In our evaluation of nine Convolutional Neural Networks with various input sizes, 100 and 400 Gbps optical interconnects reduce the training time by an average of 20.6% and 35.6%, respectively, compared to the baseline system with 25.6 Gbps electrical ones.