Researchers at Peking University have built an all-optical chip interconnect that achieves more than a 100-fold increase in AI inference speed while using roughly one-ninth the compute power of a commercial GPU, according to a study published in the journal National Science Review.
The system tackles one of the most persistent bottlenecks in AI hardware: data movement. In conventional GPU-based architectures, compute units spend much of their time waiting for data to travel across electrical interconnects, with memory transfer delays often dwarfing computation time. Peking University’s prototype replaces those electrical links with an on-chip photonic network that shuttles data at the speed of light.
The core components are a 400 Gbps silicon photonic transceiver for electrical-to-optical conversion and a custom 16×16 optical switch chip that routes data between computing nodes with an aggregate switching bandwidth of up to 6.4 Tbps. The switch achieves a total optical loss of less than 5 dB, including coupling loss, eliminating the need for external optical amplification.
In a demonstration, the team ran a five-layer convolutional neural network for image denoising, assigning each layer to a separate computing unit connected through the optical switch. Feature maps streamed directly from layer to layer through the photonic network, bypassing the memory store-and-forward delays that plague electrical interconnects.
The switch’s spectral response exceeds 100 nanometers, making it ready for wavelength-division multiplexing, a technique that could further multiply bandwidth by transmitting multiple data channels on different colors of light simultaneously.
“Specific objectives can be realized under limited computational resources when algorithms, processor micro-architectures and chip-level interconnections are co-designed,” the authors wrote.
The broader implications extend beyond raw performance. The optical fabric could alleviate unsustainable energy consumption in data centres and optimize latency or power use in edge-computing scenarios, where compute and energy budgets are sharply constrained. The team believes that co-packaged optics, faster silicon photonic transceivers, and improved AI chip interfaces could turn these “optical supernodes” into a practical foundation for distributed computing.
Sources: China’s optical network makes computing 100x faster with fewer chips (Interesting Engineering, July 2026); National Science Review paper