Researchers from the Chinese Academy of Sciences and the University of Ottawa constructed a monolithic chip integrating four optical analog cores, reporting the demonstration in Nature Communications. The high-throughput optical processing unit achieves a processing speed of 65.04 trillion operations per second and a compute density of 5.16 trillion operations per second per square millimeter.
The optical processing unit operates across 124 parallel task channels simultaneously. To achieve this throughput, the monolithic hardware combines coherent interference, wavelength-division multiplexing, and spatial parallelism on the single chip.
Neural network testing
Using the hardware platform, the team built an optoelectronic convolutional neural network that handles vision computing tasks. The system combines the chip's parallel four-kernel optical convolution and average pooling operations with electronic circuits that perform nonlinear activation and fully connected matrix operations.
In classification tests using the standard MNIST image dataset, the optoelectronic network achieved an accuracy rate of 95.08 percent. This result marks a 9.20 percent performance improvement over an equivalent single-core optical processing system.
Institutional background
The research team includes Xiangyan Meng, Junshen Li, Menghan Yang, Kangwei Fei, Yanzhen Li, Wei Li, Ninghua Zhu, Nuannuan Shi, and Ming Li. The authors hold affiliations across the State Key Laboratory of Optoelectronic Materials and Devices in Beijing, the University of Chinese Academy of Sciences, and the Microwave Photonic Research Laboratory in Ottawa.
Nature Communications received the initial manuscript on June 23, 2025, accepted it on July 6, 2026, and published the final open-access paper on July 31, 2026. Research funding for the project came from the National Natural Science Foundation of China under grant numbers 92573205, 62235011, 62505309, and 62535015, alongside support from the Beijing Nova Program and the China Postdoctoral Science Foundation.
