Chinese scientists develop optical AI chip 100x faster than Nvidia

A team from Shanghai Jiao Tong University and Tsinghua University has developed an optical computing chip, LightGen, that reportedly outperformed Nvidia’s A100 chip in speed and energy efficiency by over 100x for generative AI tasks.

LightGen uses photonic neurons—over 2 million integrated onto a single chip—to process and generate high-resolution images and videos using the speed of light rather than electrons.

The research, led by Professor Chen Yitong and published in Science, showed LightGen could perform operations such as image generation, denoising, and style transfer with images rich in detail.

LightGen uses a novel unsupervised training algorithm, removing the need for large labeled datasets and relying on statistical pattern recognition.

Experiments found the chip achieved a computing speed of 35,700 TOPS and an energy efficiency of 664 TOPS/watt, at a conservative estimate.

Researchers said these results suggest LightGen and photonic computing could help address the rising energy demands of AI.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Benchmark gaps put the 100x claim in doubt

The 100x edge over Nvidia’s A100 hinges on precision, workload choice, and energy accounting. The write-up omits whether I/O or ADC costs were counted, and it does not specify model types.
LightGen uses an unsupervised training method, while most A100 benchmarks use supervised deep learning. That split can skew a head‑to‑head comparison.
The 35,700 TOPS number likely covers optical ops native to the photonic stack, not the broad matrix work that drives many AI jobs. Performance may not carry over beyond the shown image generation tasks.
We still need the Science paper’s supplementary methods on test setup, model sizes, and datasets. Any claim that LightGen is ready for production hardware feels early.

Data centers can trial photonic inference in 2025 if vendors ship

If photonic accelerators move from labs to products, cloud and enterprise teams can line up vendors with pilots or hardware access. Lightmatter and Lightelligence build photonic AI platforms 1.
Akhetonics plans its first all‑optical processor by mid‑2025 1. Infra teams can schedule trials for power‑hungry generative inference.
AI platform leads can map vendors now, from Ayar Labs’ optical chiplets backed by Nvidia 1 to Celestial AI after Marvell’s buyout 1. That map helps time integration tests or a wait‑and‑see plan.