Lightmatter publishes a photonic processor running deep neural networks in Nature

Lightmatter

Lightmatter researchers published in Nature a multi-chip photonic processor that executed ResNet, BERT, and DeepMind-style reinforcement-learning models at near 32-bit electronic accuracy without specialized retraining. The six-chip module integrates 50 billion transistors and one million photonic components, delivering 65.5 trillion 16-bit operations per second at 78 watts electrical and 1.6 watts optical power.

• First commercially available photonic AI accelerator to run unaltered, off-the-shelf AI models (ResNet, BERT, Atari RL) without fine-tuning or quantization-aware training.
• Performance: 65.5 TFLOPS of 16-bit compute, consuming 78 W electrical and 1.6 W optical power in a 3D-packaged six-chip module.
• System combines photonic tensor cores with control dies using high-speed optical interconnects; scales far beyond prior photonic systems limited to simplified benchmarks.