Article

US startup Neurophos, based in Austin, Texas, claims to have developed a novel optical processor that offers up to ten times the computing power of Nvidia's latest AI supercomputer, Vera Rubin NVL72, while maintaining similar energy consumption. The technology centers on an Optical Processing Unit (OPU) that processes data using light.

Neurophos' Tulkas T100 accelerator operates at 56 GHz and features a single photon sensor with a matrix size of 1,000x1,000, compared to typical AI GPUs that use 256x256 matrices. While still in testing, large-scale production is anticipated for 2028, utilizing existing semiconductor processes, though challenges remain in vector processors and static memory requirements.