"Neurophos is among those trying to upend Moore's Law and make good on analog computing's long-promised yet largely untapped potential. 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."
"Neurophos CEO Patrick Bowen tells El Reg this is possible in part because of the micron-scale metamaterial optical modulators, essentially photonic transistors, that the company has spent the past several years developing. "The equivalent of the optical transistor that you get from Silicon Photonics factories today is massive. It's like 2 mm long. You just can't fit enough of them on chip in order to get a compute density that remotely competes with digital CMOS today," he explained."
Moore's Law is slowing while the energy required for generational performance gains is increasing, prompting exploration of alternative architectures. Neurophos is developing an optical processing unit (OPU) designed to deliver 470 petaFLOPS of FP4/INT4 compute, about ten times current leading GPUs while consuming similar power. The approach uses micron-scale metamaterial optical modulators that act as photonic transistors and are roughly 10,000× smaller than conventional silicon photonics optical transistors. First silicon demonstrating the technique was produced in May using a standard CMOS process. The chip centers on a single photonic tensor core of 1,000×1,000 processing elements occupying roughly 25 mm².
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]