US startup Neurophos has raised $7.2m to commercialise a silicon metamaterial for a low power optical AI chip.

The optical AI spinout from Duke University is developing a proprietary metasurface that serves as a tensor core processor enabled by its advanced optical properties. It has also joined the Silicon Catalyst accelerator and is looking to hire a team of engineers in Austin, Texas.

The company has been funded in a round led by Gates Frontier and supported by MetaVC, Mana Ventures, and others.

Using non-linear optical materials is a way of delivering machine learning inference at very low power as it relies on the design of the material to process data rather than power hungry GPUs. Other startups such as Cognifiber in Israel  are developing optical AI chips using different materials.

Optical AI processor to cut data centre power consumption
Cognifiber raises $6m for In-Fibre photonic processing

AI optical computer analyses Covid-19 X-rays

“Unfortunately, despite vast amounts of capital having recently been invested in optical compute for AI, the success of the field has been limited, largely because existing optical devices are too large and bulky to scale,” says Neurophos. “However, metamaterials enable new paradigms for controlling the flow of light. The discovery of metamaterials has unleashed an enormous burst of creativity, leading to demonstrations of invisibility cloaks, negative refractive index materials, and many other exotic products.”

Neurophos plans to use its optical metasurfaces for high-speed silicon photonics to drive an in-memory processor capable of fast, efficient, AI compute in the data centre.

Its metamaterial-based optical modulators are more than 1000 times smaller than those from a standard foundry PDK (Process Design Kit). This enables a technology roadmap to deliver over 1 million TOPS (Trillions of Operations Per Second) of performance. For comparison, an Nvidia H100 SMX5 today delivers at most 4000 TOPS of DNN (Deep Neural Network) performance.

This feeds a Compute-In-Memory (CIM) processor architecture for which is fed by high-speed silicon photonics to deliver fast, efficient matrix-matrix multiplication, which make up the overwhelming majority of all operations when running a neural network.

The metasurface-enabled optical CIM elements are thousands of times smaller than traditional silicon photonics modulators, enabling the processing of vastly larger matrices on-chip. This results in higher computational density that is hundreds of times more energy efficient than alternatives.

Demo of world’s first optical network-on-chip processor
Optical processing AI startup launches 4U accelerator

UK funds startup Lumai for optical AI 

“The most important factor in optical processors is scaling. Optical processors become both exponentially faster and more energy efficient on a per-operation basis as you make them larger. This means that in a finite chip area, the most important factor is how small you can build the optical devices that compose your processor. By leveraging metamaterials in a standard CMOS process, we have figured out how to shrink an optical processor by 8000X, which will give us orders of magnitude improvement over GPUs today,” said Patrick Bowen, CEO of Neurophos.

MetaVC Partners provided Neurophos’ initial funding and an exclusive license to the fund’s metamaterials IP portfolio for optical computing. Neurophos was spun out of Metacept, an incubator led by David R. Smith, James B. Duke Professor of Electrical and Computer Engineering, focused on creating metamaterials-based companies and collaborating with Professor Smith’s research group at Duke University. Neurophos CTO Tom Driscoll previously founded metamaterials-based radar company Echodyne.

“The Neurophos team has realized that the really fundamental problems of analog inference processing require a breakthrough at the level of the fundamental physics of the optical modulators. Their metamaterial is a ground-up breakthrough enabling an extraordinarily dense computing chip for next-generation AI applications,” said David Smith at Duke University.

“We’re delighted to be the newest Silicon Catalyst Portfolio Company. They have a proven track record in accelerating companies at our stage of development. Their deep experience and vast semiconductor industry ecosystem is invaluable as we look to reshape the AI landscape,” said Bowen.

“Neurophos represents much-needed progress in analog optical computing, bringing the performance of silicon photonics to the existing manufacturing infrastructure of CMOS foundries. We are confident that they will be one of the leaders of the next generation of AI hardware. This is how you get to tomorrow quickly and without wasted capital,” said Paul Pickering, Managing Partner at Silicon Catalyst.