Photonic nonvolatile memory for optical matrix multiplications

Abstract

Here we present a novel concept and its optimization of multi-level quantized non-volatile photonic memories based on a compact hybrid phase-change-material GSST-silicon Mach Zehnder modulator, with low insertion losses to serve as node in a photonic neural network. We demonstrate a 3-bit (8-state) photonic nonvolatile memory heterogeneously integrated into silicon PICs. We show switching operation of this device from the crystalline to the amorphous state using thermal heaters on-chip. We then show how these photonic memory elements can be utilized to design and demonstrate photonic tensor core functionality in vector matrix multiplication (VMM) engines with a compelling runtime complexity of O(1) uses O(N^3) resources (devices) could perform in the range 2-500fJ/MAC, 1-50 TMACs/mm^2, and ~100ps (1 clock cycle) per VMM operation.