Researchers at Peking University have developed a new all-optical interconnect system that links standard electronic chips, claiming to boost distributed AI inference speeds by more than 100 times while consuming only one-ninth of the typical computational power.
The system represents a significant architectural shift, moving away from traditional electronic data transfer between chips to optical methods, which could drastically reduce latency and energy consumption in large-scale AI deployments.
The breakthrough comes as China continues to narrow the technological gap with the United States in semiconductor design, despite ongoing challenges in advanced manufacturing capabilities.
Domestic chip designers are increasingly focusing on innovative architectures and specialized applications to offset limitations in access to cutting-edge fabrication tools.
This optical interconnect technology could provide a competitive edge in specific AI workloads, particularly those requiring high-speed data transfer between multiple processing units.
Previous developments from Chinese research institutions have included neuromorphic chips capable of modeling complex brain structures in real time, with performance claims significantly outpacing leading commercial GPUs.