LightGen unveils AI chip 100x faster and more efficient than Nvidia GPUs, still in lab phase

LightGen is an all-optical generative AI chip developed by teams from Shanghai Jiao Tong University and Tsinghua University, published in Science on December 18, 2025. It integrates over 2 million photonic neurons in a compact 136.5 mm² 3D-stacked design, using light instead of electricity for processing. Unlike prior photonic chips limited to basic tasks, LightGen handles complex generative workloads like high-resolution image synthesis without breaking images into patches.

​Performance vs Nvidia GPUs

LightGen delivers 35,700 TOPS at 664 TOPS/W, exceeding Nvidia's A100 GPU by over 100x in speed and energy efficiency for tasks like 512x512 image generation, style transfer, denoising, and 3D scene creation. It achieves this through an optical latent space for data compression and novel Bayes-based training, enabling probabilistic outputs rivaling Stable Diffusion. Benchmarks show comparable output quality with dramatically lower power draw, addressing AI inference's growing energy demands.

Key Technical Innovations

The chip employs metasurfaces for optical encoding/decoding, coupling full-resolution images into fiber arrays as latent representations. 3D packaging boosts density to 262 TOPS/mm², far surpassing electronic chips. Training occurs optically via unsupervised methods, reducing reliance on labeled datasets and enabling novel content generation like animal images or videos.

​

Current Limitations

LightGen remains a lab prototype, relying on bulky external lasers and specialized metasurface fabrication not yet scalable for mass production. It excels in vision-specific inference but cannot run general software, train large models, or replace CPUs/GPUs broadly. Commercialization requires standard chip factory integration, though photonic trends suggest viability as specialized AI accelerators.