Peking University researchers said their prototype distributed-AI system ran an image-denoising model more than 100 times faster than a commercial GPU while using about one-ninth of the compute resources.
The gain came from replacing electrical chip-to-chip links with an on-chip optical network that sends feature maps directly between processors, cutting memory-transfer delays and keeping compute units continuously active.
At the core are a 400 Gbps silicon photonic transceiver and a custom 16×16 optical switch chip, giving the network up to 6.4 Tbps of aggregate switching bandwidth.
The switch recorded less than 5 dB total loss and error-free transmission across multiple paths, while supporting more than 100 nm spectral response for future wavelength-division bandwidth expansion.
The team said the architecture points to a lower-power path for scaling AI in data centers and edge systems, with the study published in National Science Review.
With giants like Nvidia pushing optical tech, can this university lab win the race for next-generation AI hardware?
China's new optical chip is 100x faster. Can this breakthrough solve AI's unsustainable energy crisis?
China’s LightGen Optical AI Chip: 100x Speed, Fractional Energy Use, and the Future of Sustainable AI Hardware
Overview
As generative AI models become more advanced, the need for greater computational power and energy efficiency is pushing traditional electronic chips to their limits. In response, researchers are exploring new computing methods, with optical chips—using light instead of electricity—emerging as a promising solution. This led to a major breakthrough in late 2025 and early 2026 from China, where a team led by Professor Chen Yitong developed LightGen, an all-optical AI chip. LightGen demonstrates how light can process and generate information, offering a new path to overcome the speed and energy bottlenecks of conventional processors.