Gazelle: A Universal Photonic Computing Platform

Photonic computing has emerged as a promising alternative to electronic architectures, offering ultra-low latency, massive parallelism, and high energy efficiency. However, existing photonic computing systems are often limited to small-scale or simplified neural networks, failing to demonstrate capabilities on mainstream deep learning models. In this paper, we present ​Gazelle, a hybrid optoelectronic computing platform that successfully deploys and executes large-scale convolutional neural networks (CNNs), including the 50-layer ResNet-50, on actual photonic hardware. To the best of our knowledge, this is the ​deepest photonic-based CNN​ demonstrated to date. Gazelle achieves ​92% TOP-5 and 75% TOP-1 accuracy​ on the ImageNet-1K dataset, matching the performance of conventional electronic systems while leveraging the inherent efficiency of photonic computing. Our work bridges the gap between theoretical photonic advantages and practical AI deployments, providing a scalable and robust framework for complex deep learning tasks in optical computing.