Energy Efficiency of Photonic Convolution for Artificial Intelligence Workloads

Jonas Weiss, Pascal Stark, Lorenz K. Müller, Folkert Horst, Roger Dangel, Bert Jan Offrein

Abstract: Energy Efficiency of Artificial Intelligence (AI) workloads increasingly becomes a challenge as i) they are being adopted by a growing community of industries and ii) the AI models used are growing tremendously in complexity. We can identify certain common operations which contribute to the bulk of the computations of these workloads and thus also to the overall energy footprint. Besides data-transport and generic multiply-accumulate operations, convolutions with relatively small kernels constitute a substantial part of today’s AI workloads. In this paper we will investigate potential and limitations of optical convolutional processors for AI workloads to improve their energy efficiency. We underline our findings with a thorough system analysis and with simulation and measurement results of a sequential lattice filter type optical convolutional processor on silicon.