Cornelis Networks recently made waves with the launch of its data center networking platform based on the Omni-Path architecture that it claims outperforms competing systems like InfiniBand, a performance delta CEO Lisa Spelman said almost demands attention from potential customers.

The Cornelis Networks’ CN5000 platform is a complete system that includes network interface cards (NICs), switches, software, and cabling. The networking aspect is based on the Omni-Path architecture it gained when it was spun out of Intel in 2020, with the initial platform launch supporting 400 Gb/s (400G) networking speeds.

Cornelis Networks said its end-to-end networking platform can carry more data, faster, and with less latency than what the InfiniBand Next Data Rate (NDR) 400G platform can provide. Spelman said this networking performance can also boost compute utilization within those data center environments.

“Products that are targeted toward this space … they're getting less than 50-percent compute utilization, and that can depend on the cluster and all that,” Spelman said. “But you see as low as 25 percent, you see lots in the 35-percent range, and you see best in class at a hyperscaler with a couple-hundred performance engineers getting to that 50-percent mark, that's a tremendous amount of compute to be left on the table.”

Spelman explained that this shortfall leads to direct cost implications for customers that run deep into their data center environments.

“Yes, there's established competitors, but also, yes, there's a giant gaping market gap in how much the network is holding back compute utilization,” Spelman said. “So that's where we see the opportunity we have.”

Spelman said Cornelis’ stacked approach also provides simplified differentiation in the market, especially when it comes to day-two operations. This revolves around its use of open-source software that is compatible with platforms like Nvidia’s Compute Unified Device Architecture (CUDA) environment and AMD’s Radeon Open Compute (ROCm) platform.

“When you install our products, it is not extra application work for the actual users,” Spelman said. “We don't want it to require a bunch of hand tuning or extra work on behalf of your technical team to use. It's meant to be the easiest, smoothest, bring-up in the industry, and we're holding ourselves to that standard.”

Spelman did acknowledge that this ease of installation is a given in the market, “but we're just saying we are going to isolate you from a bunch of that work, and it's an area where we think we differentiate.”

Cornelis is initially targeting its platform at enterprise, government, and academia that are focused on AI and high-performance computing (HPC) workloads. Despite the potential complexity involved in deploying Cornelis’ full platform stack, Spelman said the performance and efficiency benefits outweigh those challenges.

“It puts us in a position where, quite frankly, we can't be ignored,” Spelman said. “If you can deliver that type of performance with no changes to your compute, literally just swapping out the network at the same bandwidth. We're not even saying, oh, take someone's 200 compared to our 400, we're saying 400 to 400, no change to your compute, you can get a 45 percent higher application performance and an even higher price performance delta, it's pretty hard to say, ‘no, I'm not going to take your call.’ It's too much return on the investment to leave on the table. It's too much TCO opportunity. So that's our way in.”

Roadmap intersection with UEC

Once in, Cornelis does have a roadmap toward the future.

The firm’s initial platform release came just days before the Ultra Ethernet Consortium (UEC) released its initial 1.0 specifications for a standards-based networking environment. Those specs include standards for congestion management, credit-based flow controls, and resiliency.

Spelman noted that Cornelis is part of that consortium and designed its initial platform release to be “compatible” with the UEC standard. However, it did not use that UEC specification due to timing.

“It wasn't in our best interest to hold product to wait for an industry consortia that can sometimes move quickly and sometimes move kind of slowly depending on how things are going,” Spelman said.

More defined compatibility will come from Cornelis’ next-generation platform that will support 800G speeds, which the vendor describes as “cross-paradigm integration.”

“What we've done is, instead of just having Ethernet and Omni-Path be like side-by-side, you choose one or the other,” Spelman said. “When you're using Ethernet mode, if you will, you actually still get access to some of those differentiated Omni-Path features. So it'll be differentiated and highly performant Ethernet in ways that the industry hasn't actually seen yet.”

This will lead to then Cornelis’ 1.6 Tb/s (1.6T) offering, which is set to integrate UEC standards with the Omni-Path technology.

“We use [the UEC specifications] as a guiding post and light but we also check everything we're doing with our customers and make sure that we're going to meet their needs in the market on time, and we intercept with the consortia as it fits in with the roadmap,” Spelman said.

Cornelis’ current roadmap also steers clear of large hyperscalers, which Spelman noted were on a different networking trajectory. This includes them being more focused on their transitions to 800G and eventually 1.6T systems.

“Our 800-gig that has that Ethernet super-NIC, that's where we start to have the opportunity to address the next wave cloud, or Neo cloud, and a hyperscale cloud market, then continuing on from there with the 1.6T. We have a tremendous amount of flexibility in that,” Spelman said. “We can do custom and semi-custom versions, and we can work with those customers to bring them these levels of performance in a lot of different form factors and structures to meet their specific needs. … So we intend to service that market with our next-generation products.”