Home / Technology / Optical Chip Shatters AI Compute Limits
Optical Chip Shatters AI Compute Limits
31 Jan
Summary
- Neurophos Tulkas T100 optical chip boasts 470 petaFLOPS compute.
- Chip integrates 768GB HBM for memory-intensive AI workloads.
- Full production expected by mid-2028 with initial thousands shipped.
Austin-based startup Neurophos is developing the Tulkas T100, an optical processing unit designed to significantly advance AI computation. Funded by Bill Gates' Frontier Fund, the chip aims for 470 petaFLOPS of FP4 and INT4 compute within a 1-2kW power envelope.
Its optical tensor core is substantially larger than conventional AI GPUs, enabling higher compute density. Neurophos leverages optical transistors, claimed to be 10,000 times smaller than current silicon photonics components, to potentially extend Moore's Law.
The Tulkas T100 operates at a high 56GHz clock rate. It features a dual reticle design with 768GB of HBM to support demanding AI workloads, focusing initially on token processing for large language models.
