Neurophos Secures Impressive $7.2m Funding Boost

Neurophos, a spinout from Duke University and Metacept, has recently secured a $7.2 million seed round to advance its work in metamaterials and optical AI inference chips.

The funding round was led by Gates Frontier, with support from MetaVC, Mana Ventures, and other investors. This financial boost will allow Neurophos to develop a proprietary metasurface, which serves as a tensor core processor, leveraging its advanced optical properties.

"By utilizing metamaterials in a standard CMOS process, we have successfully reduced the size of an optical processor by 8000 times. This breakthrough will provide us with significant improvements over the current capabilities of GPUs," explains Neurophos CEO Patrick Bowen.

Neurophos combines two significant breakthroughs in its technology. The first is an optical metasurface that enables silicon photonic computing, capable of ultra-fast AI inference that surpasses the density and performance of traditional silicon computing and silicon photonics.

The second breakthrough is a Compute-In-Memory (CIM) processor architecture, which is powered by high-speed silicon photonics. This architecture enables fast and efficient matrix-matrix multiplication, which is a crucial operation in running neural networks.

While GPUs have been successful in accelerating AI workloads, their digital approaches are often limited by power consumption. Photonics, on the other hand, can significantly reduce power consumption. However, existing optical devices are too large and bulky to scale effectively. This is where metamaterials come into play, as they enable new possibilities for controlling the flow of light.

Neurophos' optical metasurfaces are specifically designed for use in data centers. The company plans to utilize high-speed silicon photonics to drive a metasurface in-memory processor, capable of delivering fast and efficient AI compute.

Compared to standard foundry PDK, Neurophos' metamaterial-based optical modulators are more than 1000 times smaller. This allows for a technology roadmap that can deliver over 1 million TOPS (Trillions of Operations Per Second) of performance. In comparison, an Nvidia H100 SMX5 currently delivers a maximum of 4000 TOPS of DNN (Deep Neural Network) performance.

The metasurface-enabled optical CIM elements developed by Neurophos are thousands of times smaller than traditional silicon photonics modulators. This breakthrough enables the processing of significantly larger matrices on-chip, resulting in an unprecedented increase in computational density. In optical computing, energy efficiency is directly proportional to array size, making Neurophos' processor hundreds of times more energy efficient than alternative solutions.