Optical interconnect startup Xscape Photonics has raised $37 million in funding in what it described as an extension round, bringing its total Series A investment to $81 million.

The round doubles its valuation, and comes alongside the launch of a new device called FalconX that enables ultra-fast, high-capacity data transmission for artificial intelligence data centers. Today’s round was led by Addition and saw additional commitments from existing investors including IAG Capital Partners and Nvidia Corp., the company said.

Xscape is building a new kind of interconnect that’s designed to link thousands of graphics processing units and other AI processors that power AI workloads. It says that data centers need a better way to connect those chips, as existing interconnects are limited by their available bandwidth, causing a bottleneck for AI.

Its interconnects are based on silicon photonics, which use lasers to transmit data instead of electrical signals sent over copper wires, allowing for much greater bandwidth. They can transmit data much faster than traditional interconnects, accelerating the performance of GPUs.

FalconX is the company’s first product. It’s described as the industry’s first fully redundant external laser small form-factor pluggable device of its kind, and it’s capable of emitting up to eight wavelengths and colors of light at once. Doing this, it can send multiple data streams simultaneously across the same link, dramatically increasing the volume of data it’s able to transmit, the company said.

According to Xscape co-founder and Chief Executive Vivek Raghunathan, existing interconnects have become the major bottleneck in AI performance. The copper wires they use can only handle so much data, but modern chips process information at much faster speeds. As a result, they’re often left sitting idle, waiting for data to arrive, so they can process it. “Rapidly increasing bandwidth, power and cost demands of AI workloads have created a critical hardware bottleneck, forcing developers to use just a fraction of their GPUs’ capacity, thereby limiting the revolutionary potential of AI,” Raghunathan said.

FalconX overcomes this by generating eight colours from a single laser module to deliver multi-terabits-per-second of data bandwidth. “FalconX represents the opening of a brand-new data highway, on which eight lanes of data traffic all travel reliably at light speed, allowing the entire data center to function as one giant GPU,” Raghunathan continued.

It’s a good start, but Raghunathan says he has ambitions to increase the bandwidth of photonic-based interconnects dramatically. He intends to scale the company’s platform to handle even more wavelengths of light.

The first step will be to develop an interconnect that can handle 16 wavelengths, followed by 32 and 64, before ultimately reaching 128 colors. The company has already developed a prototype device called CombX that can transmit data in 16 wavelengths, and has successfully demonstrated that capability through a collaboration with Tower Semiconductor Inc. So that’s the next product to look out for.

Addition founder Lee Fixel believes advanced photonics are quickly becoming critical to AI infrastructure. “With innovations like FalconX’s eight-wavelength redundant laser and the ChromX roadmap to 128-plus colors, Xscape is addressing the bandwidth bottlenecks that constrain AI cluster performance,” he said.

Xscape faces a lot of competition in the photonics sector, with technology giants like Intel Corp., Hewlett Packard Enterprise Co. and IBM Corp. focused on developing their own laser-based interconnects. Marvel Technology Inc., which recently acquired Celestial AI, is another potential rival. There are plenty of startups developing alternatives too, such as the Swiss company Lightium AG and Luminous Computing Inc.

Image: Xscape Photonics