Peking University's optical chip system boosts AI speed 100x with 90% less compute

Peking University researchers have unveiled an all-optical interconnect system that dramatically accelerates AI inference. By linking standard field-programmable gate array (FPGA) chips with custom-designed optical communication hardware, the system boosts distributed inference speeds by over 100x while consuming only one-ninth of the compute resources typically required. The key components are two silicon photonic transceiver chips running at 400 Gbps each, converting electrical signals to optical and back. This work, published in National Science Review and led by Shu Haowen and Wang Xingjun, challenges the brute-force approach of adding more GPUs.

The system leverages FPGAs as "Lego" building blocks — programmable devices already used in high-parallelism tasks like autonomous driving and data centers. The optical interconnects act as high-speed "joints" that enable efficient data transfer between chips. Instead of scaling up GPU clusters and data centers, this method suggests a more efficient path by optimizing interconnect technology. The breakthrough could significantly reduce energy consumption and hardware costs for AI workloads, potentially reshaping how distributed computing systems are designed. With AI's insatiable demand for compute, this optical approach offers a scalable, energy-efficient alternative.