Silicon Micro-Disk Resonator Crossbar Array for High-Speed and High-Density Photonic Convolution Processing

TL;DR

This paper presents a silicon photonic micro-disk resonator crossbar array that significantly enhances the speed and density of optical matrix-vector multiplication for AI applications, demonstrating high accuracy and scalability.

Contribution

Introduction of a novel MDR crossbar processor with dual MDRs at each crosspoint for improved density and speed in optical MVM operations.

Findings

Achieves 96% accuracy on MNIST dataset.

Supports up to 160 TOPS processing speed.

Reaches a computational density of 25.6 TOPS/mm2.

Abstract

Advanced artificial intelligence (AI) algorithms, particularly those based on artificial neural networks, have garnered significant attention for their potential applications in areas such as image recognition and natural language processing. Notably, neural networks make heavy use of matrix-vector multiplication (MVM) operations, causing substantial computing burden on existing electronic computing systems. Optical computing has attracted considerable attention that can perform optical-domain MVM at an ultra-high speed. In this paper, we introduce a novel silicon photonic micro-disk resonator (MDR) crossbar signal processor designed to support matrix-vector multiplication (MVM) with both high processing speed and enhanced computational density. The key innovation of the proposed MDR crossbar processor is the placement of two MDRs at each crosspoint, enabling simultaneous routing and weighting functions. This design effectively doubles the computational density, improving overall performance. We fabricate a silicon photonic MDR crossbar processor, which is employed to perform convolutional tasks in a convolutional neural network (CNN). The experimental results demonstrate that the photonic processor achieves a classification accuracy of 96% on the MNIST dataset. Additionally, it is capable of scaling to a computational speed of up to 160 tera-operations per second (TOPS) and a computational density as high as 25.6 TOPS/mm2. Our approach holds significant promise for enabling highly efficient, scalable on-chip optical computing, with broad potential applications in AI and beyond.