As the pace of Moore’s Law slows, chip designers are exploring alternative architectures. Neurophos, a startup backed by Bill Gates, is developing an innovative optical processing unit (OPU) that promises to significantly enhance computational performance.
What is the Optical Processing Unit?
Neurophos’s OPU is designed to achieve a remarkable 470 petaFLOPS of FP4/INT4 compute, which is approximately ten times the performance of Nvidia’s latest Rubin GPUs, while consuming a similar amount of power. This capability stems from the company’s development of micron-scale metamaterial optical modulators, which function as photonic transistors.
How Does it Work?
According to Neurophos CEO Patrick Bowen, traditional optical transistors produced by Silicon Photonics factories are significantly larger, measuring around 2 mm. In contrast, Neurophos’s optical transistors are about 10,000 times smaller, allowing for greater compute density on a chip. The first silicon prototype, developed in May, demonstrated compatibility with standard CMOS processes, facilitating integration with existing manufacturing technologies.
Neurophos’s OPU features a unique photonic tensor core, which is significantly larger than those found in most AI accelerators. This tensor core measures 1,000 by 1,000 processing elements and is designed to operate at a frequency of 56 GHz. Unlike conventional GPUs that require multiple tensor cores, Neurophos’s architecture only needs one, occupying approximately 25 mm² of the chip area.
Performance and Future Plans
The upcoming OPU, codenamed the Tulkas T100, will utilize a dual reticle design and include 768 GB of HBM. It aims to deliver 470 petaOPS while consuming between 1 to 2 kilowatts of power under load. However, these performance metrics are still in the developmental phase, with full production not expected until mid-2028. Bowen anticipates that the initial production will be limited to thousands of chips, not tens of thousands.
Neurophos envisions the Tulkas T100 primarily serving as a prefill processor for AI workloads, similar to Nvidia’s Rubin CPX. The company is also working on a proof of concept chip to validate its performance claims, supported by a recent $110 million Series-A funding round led by Gates Frontier.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
