Deep neural network inference on an integrated, reconfigurable photonic tensor processor

Abstract

Abstract Artificial neural networks set the pace in machine vision, natural language processing, and scientific discovery, but their performance depends on fast and efficient tensor computations. Analog photonic systems are a promising alternative to digital electronics because they enable ultra-fast, low-latency computing while avoiding capacitive charging losses and electrical crosstalk. Here we present a photonic tensor processor for deep neural network inference, integrated into a standard 19-inch rack unit with a high-speed electronic interface to PyTorch for seamless hardware deployment. The processor implements an all-optical crossbar with nine inputs and three outputs for parallel intensity-based accumulation of weighted signals. Fabricated in imec’s iSiPP50G silicon photonics platform, the chip integrates electro-absorption modulators and photodiodes for scalability and compatibility with high-volume manufacturing. An integrated self-injection-locked microcomb provides stable multi-wavelength carriers. We demonstrate inference on MNIST and CIFAR-10 with 98.1% and 72.0% accuracy, highlighting a compact, reprogrammable platform toward scalable high-speed optical AI accelerators.