Silicon photonics startup nEye raises $58M to light up AI data centers – News

Optical networking chip startup nEye Systems Inc. wants to become the nerve center for artificial intelligence data centers after closing on $58 million in funding from a host of big tech heavyweights.

Today’s Series B round was led by CapitalG, a growth-stage fund backed by Google LLC’s parent company Alphabet Inc., and saw participation from Microsoft Corp.’s M12, plus Micron Ventures, Nvidia Corp. and Socratic Partners. It brings the company’s total amount raised to more than $72 million, following a previous, undisclosed seed funding round.

The Emeryville, California-based startup is developing a networking chip that utilizes optical technology to transmit data between AI chips as a form of light, as opposed to electrical signals. Its technology is hugely interesting to the likes of Google and Nvidia, because it can facilitate faster chip-to-chip communications at much lower cost than existing network interconnects.

Most modern data centers today rely on older interconnects that are based on electrical switches, but these have significant bandwidth limitations and consume lots of energy. This causes a bottleneck for AI workloads that rely on huge clusters of connected graphics processing units, as they’re unable to talk to each other quickly enough. It also means higher costs due to the excessive power consumption.

The company’s wafer-scale optical circuit switch is claimed to be a more efficient and cost-effective alternative, as its direct optical connections support data transfers with virtually unlimited bandwidth. It says its chips are about 100 times smaller, 1,000 times more energy efficient, 10,000 times faster and 10 times less expensive than existing data center interconnects.

This isn’t the only startup building optical interconnects. The technology sounds very similar to what is being built by other well-funded companies, such as Lightmatter Inc., Celestial AI Inc., Ayar Labs Inc., Xscape Photonics and Lightium AG. Traditional chip giants, including Intel Corp. and IBM Corp., are also working on silicon photonics, which is another word used to describe the same concept.

However, nEye appears to have an ace up its sleeve, for it’s also developing an intelligent optical circuit switch that adapts the mesh of chip-to-chip connections within data centers on the fly. This technology can increase data center performance by establishing the most optimal connection between different chips based on the software that’s being run.

It’s an idea that Alphabet has already experimented with internally. Several years ago, Google built an AI supercomputer that it said was much more powerful than Nvidia’s most advanced offerings at the time. However, though it uses it internally — and touted it today at its Cloud Next conference in Las Vegas as an energy saver — Google has never commercialized this technology.

That’s exactly what nEye wants to do, building on the concept to create its own intelligent optical circuit switches. It promises vast improvements in data center efficiency, both for AI and other computing workloads.

“Google is a pioneer, it led the way,” nEye co-founder Ming Wu, a professor at the University of California Berkeley, told Reuters. “Other AI companies, other hyperscaler AI data center operators, they will be looking to acquire this technology rather than develop it themselves.”

The startup says it has already built prototypes of its chips and is aiming to be able to send samples of production chips to customers next year. However, it has not said when it will be able to ship them in larger volumes.

Although it’s primarily targeting the AI industry, nEye says its chips can be just as beneficial for traditional data center workloads, which also suffer from high energy costs.

CapitalG General Partner James Luo said nEye is addressing critical bottlenecks in both AI and traditional high-performance computing. “The beauty of it is that it’s applicable to both models,” he stressed.

Image: News/Microsoft Designer