Peking University links chips optically to speed AI inference

Researchers at Peking University have developed an all-optical interconnect system for linking standard electronic chips, increasing AI distributed inference speeds by over 100 times while using just one-ninth of typical computational resources. The work points to an alternative to adding more GPUs and expanding data centres as AI models drive higher demand for computing power.

The study, published in National Science Review, lists Shu Haowen and Wang Xingjun from Peking University among its corresponding authors. The system uses field-programmable gate array chips, or FPGAs, as programmable building blocks suited to highly parallel workloads in areas including missile guidance, autonomous driving and data centres.

Custom communication hardware forms the connections between the FPGA chips. One component is a silicon photonic transceiver chip running at 400 gigabits per second, converting electrical signals to optical signals and back again.