Researchers design advanced photonic chip to unlock scalable AI computing

Researchers design advanced photonic chip to unlock scalable AI computing
by Riko Seibo
Tokyo, Japan (SPX) Oct 24, 2025

A team from Huazhong University of Science and Technology and The Chinese University of Hong Kong has developed a novel deep optical neural network chip that addresses critical scaling limitations in photonic artificial intelligence systems. Existing on-chip photonic neural networks (ONNs) struggle with depth constraints due to weak cascadability in optical nonlinear activation and are limited by input size because of strict requirements on coherent light sources.

The new chip, called a partially coherent deep optical neural network (PDONN), utilizes an on-chip nonlinear activation function based on opto-electro-opto conversion, providing positive net gain and enabling the stacking of deeper network layers. Its architecture includes convolutional layers for rapid dimensionality reduction, helping to process larger input images.

Unlike conventional designs that depend on expensive narrow-linewidth lasers, the PDONN employs partially coherent light sources, such as LEDs or amplified spontaneous emission sources. This substitution significantly simplifies the system and makes scaling up the optical matrix more feasible. The PDONN integrates hundreds of optical devices within a compact 17 mm2 area, including a 64-unit input layer, two convolutional layers, and two fully connected layers-the largest input size and optical layer count reported for on-chip ONNs so far.

In trial runs, the chip accomplished four-class handwritten digit classification and two-class fashion image categorization with accuracies of 94 percent and 96 percent. Performance remained robust when using partially coherent light, underscoring the reliability of the architecture. The device achieved a single-inference latency of 4.1 ns and energy efficiency of 121.7 pJ/OP, highlighting its promise for low-latency, energy-efficient AI applications.

"This work demonstrates that through architectural innovation, we can use more accessible technology to build large-scale, high-performance optical computing systems," stated the researchers. "It opens up new avenues for the next generation of energy-efficient, widely applicable AI hardware".

Further improvements in modulator extinction ratio and reduced system latency are planned, potentially enhancing scalability and performance. The PDONN's innovation signals a significant milestone in optical computing for artificial intelligence.

Research Report:Scaling up for end-to-end on-chip photonic neural network inference