Scale-up interconnect startup CScale launches with $188 million in funding

CScale Inc., a startup that develops optical interconnects for data centers, started today with $188 million in funding.

The company has raised approximately three-quarters of capital through a recently completed Series C round. It was co-led by Atreides Management, Valor Equity Partners and Premji Invest. The funds were joined by several other backers, including Nvidia Corp. and Intel Capital.

CScale was founded in 2023 by a well-known electrical engineer Sanjai Kohli. Previously, he founded SiRF Technologies Inc., an early provider of GPS chips for vehicle navigation systems. Kohli is CScale’s chief technology officer, while the company’s CEO is long-time data center executive Martin Lund. Previously, Lund led Cisco Systems Inc.’s chip engineering group and Broadcom Inc.’s switch business.

CScale’s interconnection technology is intended to connect the accelerators together, forming an artificial intelligence cluster. According to the company, its hardware is optimized to run large cluster configurations. Such setups can provide better performance than current designs.

An AI data center is divided into infrastructure units called scale-up domains. Today, a scale-up domain is typically a single rack of servers. The chips in the rack practically function like a large computer. To add more processing capacity, companies must deploy additional scale-up domains.

Data moves slower between different scale-up domains than between accelerators in the same scale-up domain. In March, Nvidia announced that it was working to resolve this bottleneck. The chipmaker is developing hardware that allows multiple compact scale-up domains to be replaced by one large scale-up domain. Such a setup enables faster transfer of data between accelerators, which should improve the performance of AI workloads.

The CScale connection is intended to facilitate the development of large scale-up clusters. The company says its technology can connect thousands of accelerators in dozens of racks. CScale will transfer data between chips over optical connections that are significantly faster than traditional copper cables.

A malfunction in one of the optical links that powers a scale-up cluster can disrupt many other components. CScale says its connection can “contain” such failures, increasing cluster availability. The company has not specified how it achieves this increase in reliability.

According to a Job advertisementCScale plans to deliver its technology as part of a system-on-chip that will have an integrated central processing unit. The CPU may be required to run network automation software. CScale describes its product as an “integrated light engine.”

“As scale-up systems move toward gigawatt-class deployments, interconnect bandwidth means nothing without system reliability,” said Gavin Baker, managing partner and chief investment officer of Atreides Management. “Every optical failure is a computational error. At this scale, failures are not hypothetical – they are inevitable. CScale is developing reliable scale-up optics for next-generation AI factories.”

CScale will use its new funding to improve its technology and accelerate commercialization initiatives.

Photo: Unsplash