China’s Peking University Achieves 149X AI Speedup with Light-Linked Chips

Researchers at Peking University in China developed a new chip system that accelerates artificial intelligence processing by linking standard electronic chips with an all-optical network. The team replaced traditional copper wiring with an optical network, allowing data to flow between processors much faster. This system paired five field-programmable gate arrays, or FPGAs, with a custom optical switch to process images from the Fashion-MNIST dataset in 105.16 microseconds.

While the FPGAs delivered only 11.6% of the computational capacity of a commercially available GPU, they completed the task nearly 149 times faster. The design bypasses the typical “memory wall” bottleneck found in GPUs by sending data directly from one chip layer to the next, keeping all processors consistently busy.

The researchers utilized standard silicon fabrication techniques, a strategic advantage given U.S. export controls on advanced lithography. Though the current demonstration used a limited workload, the team aims to scale the design to 16 processors and improve input/output speeds for broader applications.