SHANGHAI, December 18, 2025 – Scientists from Tsinghua University and Shanghai Jiao Tong University have created LightGen, a photonic computing chip that achieves over 100 times greater speed and energy efficiency than Nvidia’s A100 GPU in specific generative AI tasks, according to a study published in Science .
The chip packs more than 2 million photonic neurons into a 136.5 mm² die, using laser pulses for data processing to minimize power consumption compared to traditional electronic systems.
Innovations include an “optical latent space” for compressing data and an unsupervised learning method that detects patterns without relying on large labeled datasets.
Benchmarks reported a system throughput of 3.57 × 10⁴ tera operations per second and efficiency of 6.64 × 10² TOPS per watt.
End-to-end tests showed LightGen producing comparable results to the A100 in image synthesis, denoising, style transfer, 3D editing, and video generation, while consuming far less time and energy.
Lead researcher Chen Yitong from Shanghai Jiao Tong University explained that earlier photonic designs faltered on complex generative workloads due to structural constraints. LightGen overcomes these, offering room for further scaling.
The work advances China’s efforts in alternative computing architectures amid U.S. restrictions on high-end Nvidia hardware.
Photonic systems could alleviate the growing energy demands of AI training, though moving from prototypes to commercial production presents ongoing hurdles.
