Photonic Convolution Accelerator Based on Electroabsorption-Modulated Multiwavelength DFB Laser

Abstract

Given the challenges in integrating noncoherent optical convolution computing architectures, we have proposed and demonstrated a compact and efficient photonic convolutional accelerator (PCA) based on a monolithically integrated modulated multiwavelength source. A tiny footprint of 0.45 mm² and a high compute density of 0.36 TOPS/mm² are achieved by processing real-time recognition tasks using the MNIST handwritten digits database. In our scheme, except for delay lines and WDM devices that are not suitable for implementation on the InP-based platform, all other necessary devices are monolithically integrated on a chip. These include a uniformly spaced multiwavelength source (laser arrays based on REC technology), weight-loaded units (electrically power-tunable DFB lasers), modulators for signal input (electroabsorption modulated, EAM), and semiconductor optical amplifiers (SOA) for loss compensation. This study provides a compact and efficient InP-based solution for photonic convolutional neural networks (PCNNs).