The GPUs and other chips used to train AI communicate with each other inside datacenters through “interconnects.” But those interconnects have limited bandwidth, which limits AI training performance. A 2022 survey found that AI developers typically struggle to use more than 25% of a GPU’s capacity.
One solution could be new interconnects with much higher bandwidth, according to Vivek Raghunathan, the CEO and co-founder of startup Xscape Photonics. The secret sauce, he says, is silicon photonics: silicon-based material that manipulates light to transmit data.
“Xscape has created a platform that connects various computing elements in a sustainable way, while offering the highest possible performance,” Raghunathan told TechCrunch in an interview. “The core of this platform’s scaling relies on energy-efficient, cost-effective systems that do not exist in the industry yet.”
Xscape is based in Santa Clara in the heart of Silicon Valley, but has its roots in a Columbia University lab, where three professors — Alexander Gaeta, Keren Bergman, and Michal Lipson — invented a technique they believed could be used to transmit terabytes of data over light.
The trio spun off Xscape in 2022 after recruiting Raghunathan and Yoshitomo Okawachi, a laser engineer and a longtime colleague of Gaeta’s. Raghunathan joined by way of Broadcom, where he helped found the silicon photonics team, and Intel, where he was manager for the company’s silicon photonics products.
Traditional interconnects consist of metal wires that transmit data in the form of electrical signals.
Metal-based interconnects require a lot of power — and generate lots of heat. They’re bandwidth-constrained by their medium’s conductivity. And, in datacenters with fiber-optic links between components, the interconnects’ electrical data must be converted into optical and back again, introducing latency.
Silicon photonics like Xscape’s, in contrast, draw minimal power and produce negligible heat.
“In the past, we primarily used optical communications for long-haul fiber-optic systems,” Raghunathan said. “But recent advancements are enabling the integration of optics-on-chip — in the form of silicon photonics — and bringing the optical interface from the electronic plane to the optical plane all the way into the chip.”
Xscape’s first product is a programmable laser to power datacenter fiber-optic interconnects, specifically the links among GPUs, AI chips, and memory hardware. The laser can leverage different colors of light (i.e. wavelengths) to transmit multiple data streams along the same link without interference, Raghunathan claims.
“Electrical systems densely packed together tend to produce crosstalk, interference, and other challenges,” he said. “However, within the optical domain, data can be modulated on different colors, wavelengths, or channels, and all co-propagate within the same wire or fiber — and not interfere with each other.”
Assuming the tech works as advertised, Xscape faces the same challenge as most hardware startups: manufacturing and selling its products at scale. In a possible leg up over photonics rivals like Ayar Labs and Celestial AI, Xscape’s lasers can be fabricated using the same facilities used to make the microelectronics in phones and laptops.
The first-generation laser can only emit between 4 and 16 colors. However, Xscape is already planning improved versions that’ll be able to emit up to 128.
Xscape says that it’s “actively engaged” with ten customers for potential deployments, ranging from vendors to hyperscalers — and that it has secured funding from Cisco and Nvidia, whose venture arms invested in its recent $44 million Series A round. The investments aren’t strategic, meaning that the companies aren’t currently customers. But Raghunathan notes that Cisco is one of the largest sellers of optical networking components in the world.
“This reflects Cisco and Nvidia’s trust in the value we bring to this ecosystem,” Raghunathan said.
The latest funding round was led by IAG Capital Partners, and brought the company’s total raised to $57 million. Raghunathan says that the proceeds will be put toward growing Xscape’s 24-person team, and scaling up fabrication of its lasers and related photonics tech.
“The funding will allow Xscape to push the boundaries of our platform and integrate it with simulation, high-performance compute, and AI software to help customers in all industries take their innovations to new heights,” Raghunathan said.
Xscape certainly has its work cut out for it. Aside from Ayar and Celestial, the firm competes with Intel in the multi-billion-dollar silicon photonics market. Intel claims to have shipped over 8 billion photonics chips and 3.2 million on-chip lasers since 2016.