Neurophos bets on optical transistors to bend Moore’s Law

TLDR

Neurophos, an AI chip startup, is developing an optical processing unit (OPU) capable of delivering 470 petaFLOPS of compute using micron-scale metamaterial optical modulators, or optical transistors, which are 10,000 times smaller than current alternatives. The OPU, codenamed Tulkas T100, is designed to handle AI inference workloads, particularly the compute-intensive prefill stage, and is expected to consume 1 to 2 kilowatts of power under load. While the chip is still in development, with full production not expected until mid-2028, Neurophos aims to tackle the decode phase of AI inference in the future.

The Austin, Texas-based AI chip startup says it's developing an optical processing unit (OPU) that in theory is capable of delivering 470 petaFLOPS of FP4 / INT4 compute — about 10x that of Nvidia's newly unveiled Rubin GPUs — while using roughly the same amount of power.