A new photonic processor performs neural network matrix multiplications using light interference in silicon waveguides, encoding data across optical wavelengths to bypass electron transport bottlenecks entirely. Claiming 100x inference speed with watts-scale power versus GPU kilowatts, it handles transformer models at cluster-equivalent throughput on a single chip. This emerges alongside NVIDIA’s 400 Tb/s Spectrum-X photonics Ethernet for AI fabrics, suggesting hybrid electro-optical compute as the post-electronic scaling path for edge-to-cloud inference where thermal walls cap GPU density.
Source: ConnectingAI
Link:
Leave a Reply