Photonic Chips Finally Scale Up to Power Massive AI Models

In the quest for faster, greener artificial intelligence, photonic computing—using light instead of electricity—has long been a promising contender. However, these systems face a significant hurdle: they are analogue by nature. Unlike digital systems, analogue signals suffer from noise that accumulates rapidly as data passes through the layers of a neural network. Until now, this 'error accumulation' restricted optical chips to a depth of only about ten layers, rendering them too shallow for modern, complex AI tasks.

A new study has identified the root cause of this instability: ‘propagation redundancies’. By introducing precise perturbations on the chip to decouple these computational correlations, researchers have effectively eliminated the redundancy. This led to the development of the single-layer photonic computing (SLiM) chip, a device that is robust against errors and can extend spatial depth from the millimetre scale to hundreds of metres.

The performance implications are vast. The team experimentally constructed a 100-layer neural network for image classification and even successfully ran Large Language Models (LLMs) with up to 640 layers for image generation. Operating at a 10-GHz data rate, these deep optical networks achieved accuracy comparable to ideal digital simulations. This breakthrough suggests that energy-efficient analogue hardware is finally ready to handle the state-of-the-art deep learning models that drive today's AI era.