Modeling the Energy Efficiency of GEMM using Optical Random Access Memory

Abstract

General matrix-matrix multiplication (GEMM) is the key computation kernel in many applications. GEMM has been supported on various hardware platforms, including CPU, GPU, FPGA. To optimize the performance of GEMM, developers use on-chip electrical static random access memory (E-SRAM) to exploit the data locality of GEMM. However, intensively accessing E-SRAM for GEMM can lead to significant energy consumption, which is not energy-efficient for commercial data centers. In this paper, we evaluate the optical static random access memory (O-SRAM) for GEMM. O-SRAM is a promising tech-nology that has extremely low access latency and low energy consumption compared with the traditional E-SRAM. First, we propose an O-SRAM based wafer-scale system for GEMM and a baseline E-SRAM based system. Second, we build the theoretical performance models of the two systems to analyze their energy consumption of on-chip memory accesses. Then, we conduct simulation-based experiments to evaluate the energy consumption of the two system. The evaluation results show that O-SRAM based system is 7 x more energy efficient than the baseline E-SRAM based system.