Cornelis brings Intel Omni-Path networking tech back from dead

More than a decade after it was invented and then abandoned, a network interconnect technology called Omni-Path, developed by Intel, is making a comeback. Startup Cornelis Networks, which acquired the technology from Intel in 2020, has rebuilt the product into a modern architecture designed for large-scale AI workloads.

The company on Tuesday introduced the CN5000 product family, which includes network interface cards, switches, cabling, and software. It is an alternative to Ethernet and InfiniBand, which are the most dominant networking interconnects today.

The CN5000 family is a 400G product, which is a four times bandwidth jump over the original legacy 100G product. The company plans a quick jump to 800G next year, and a 1.6T product in 2027.

“It’s a combination of enhancements of the existing features that were already there… and pulled forward capabilities like collective performance that are more important for AI,” said Lisa Spelman, CEO of Cornelis Networks.

The company is targeting enterprises, high performance computing (HPC) centers, academia, government, and cloud providers looking to create their own data centers and bring workloads on-premises. Many companies are looking to build their own AI infrastructure for cost and security reasons, Spelman said.

“A lot of enterprises we see are planning to do their inference and they’re fine-tuning on-prem simply because of the cost economics of a fully utilized cloud model,” Spelman said.

The networking gear can plug into a wide range of CPUs, AI accelerators, and GPUs that include chips from Nvidia, Intel, and AMD. The goal is to drive up compute utilization and lower cost.

“You can do more with less… within my budget, I can now tackle a bigger problem, I can train a larger model, or deliver that on a smaller budget,” Spelman said.

From Top 500 to AI

In the mid-2010s, Intel made Omni-Path the centerpiece of its strategy to put its hardware in the world’s fastest supercomputers.

In 2016, the interconnect gained momentum by appearing in 28 of the world’s fastest computers and peaked at 51 systems in 2020, after which its adoption dwindled. Intel stopped developing Omni-Path technologies in 2019, after which it was bought by Cornelis.

“We used that time over the last four years to drive … conversations and requirements with not only the existing HPC customer base, but also the AI customer base,” Spelman said.

Cornelis reworked Omni-Path’s silicon, interface, and workload management, with enhancements targeted at AI and high-performance computing. The focus is on better load-balancing and utilization rates to keep hardware occupied, which translates to better performance per watt for customers.

The idea of introducing a new networking interconnect to challenge Ethernet and Nvidia’s InfiniBand, which dominate the world’s largest AI installations, isn’t fazing Spelman.

The company claims some performance benefits for Omni-Path over other networking technologies with the same bandwidth. Optimizations of the software and hardware stacks deliver 35% lower latency, 30% higher application performance, and two times higher message rates.

“A dollar invested in the network will always have a higher payoff than another dollar invested into a GPU that’s going to sit underutilized,” Spelman said.

Analysts remain dubious

Analysts, however, have their doubts that Omni-Path can capture the industry’s interest.

“It’s hard to see Omni-Path gaining much market share, if Cornelis is targeting AI specifically,” said Jack Gold, principal analyst at J. Gold Associates. “There is room for high speed Ethernet switching, which is coming, to compete with InfiniBand, especially at a lower cost, but having another proprietary solution at this point is probably not a mass market capability,” he said, noting that Omni-Path will likely be a niche product that takes minor market share.

Dylan Patel, founder of SemiAnalysis, a semiconductor research firm, agreed. “Omni-Path technology is not yet proven for AI, which creates an adoption challenge for enterprises.”

However, Spelman argued, “When you deliver performance that customers just can’t turn away from at a great economic value, that’s how you create opportunity to break in even in a market that has strong incumbents.”