Cornelis Mounts Scale-Up Challenge To Nvidia, Teases Potential Qualcomm Tie-Up

The Intel spin-off opens a new front in its challenge to Nvidia with its entry into scale-up networking, which is meant to give companies building rack-scale AI computing platforms an open alternative to the latter company’s proprietary NVLink interconnect.

Intel spin-off Cornelis Networks is expanding beyond its scale-out networking business to challenge Nvidia in an area that is critical to rack-scale AI computing and is helping the larger vendor gain support from AI chip rivals: scale-up networking.

The Wayne, Pa.-based startup announced on Monday that its scale-up move has the potential backing of Qualcomm and that it has also raised $205 million in new funding from investors. The fresh capital will support the development of future scale-up and scale-out products as well as manufacturing and customer deployments for its new CN6000 family of scale-out products.

While Cornelis didn’t share any details about what it called a “collaboration” with Qualcomm, leaders from both companies signaled the potential to integrate their technologies together. Qualcomm plans to debut its first rack-scale platform this year with its AI200 accelerator chip, with successor products to come out in the following years.

“[Cornelis’ networking strategy] plays into a lot with what the Qualcomm team is doing on the accelerator and CPU compute side,” Cornelis CEO Lisa Spelman (pictured) told CRN in an interview. “We see a lot of shared vision on the challenges in AI infrastructure, challenges customers are facing, and a lot of opportunity and potential for our technologies to knit together and address those challenges for customers.”

Tony Pialis, general manager of Qualcomm’s data center business, said Cornelis’ “vision for an open, programmable fabric aligns” with the need to improve utilization and AI economics with a “more integrated approach across compute, memory and networking.”

“Giving customers more choice and flexibility across the infrastructure stack will be critical as AI moves toward rack-scale architecture,” he said in a statement.

Pialis is expected to join Spelman for her keynote at the AI Infra Summit this week.

Cornelis Scale-Up Products Could Give Channel More Options

Unlike Cornelis’ traditional scale-out networking business, which is focused on high-speed interconnects for servers across large clusters, the company’s foray into scale-up networking, which enables high-speed connections between processors inside rack-scale systems, may not result in channel-ready products directly from the startup.

Instead, Cornelis will use its newly revealed Active Compute Fabric architecture to give firms building rack-scale computing platforms like Qualcomm an alternative to Nvidia’s proprietary NVLink interconnect technology. This could over time give solution providers more choices in the kinds of rack-scale platforms and configurations they can offer to customers.

“I think there’s a world in which some of these channel players become mechanisms for rack delivery and servicing of that enterprise,” Spelman said.

Qualcomm has not yet announced a partner program for its data center business, though a tech analyst recently told CRN this could eventually happen as enterprise AI needs grow.

An executive at a large U.S. systems integrator told CRN that many of his enterprise customers would welcome new data center suppliers who embrace open standards because it could potentially help them with product choice, supply and pricing.

“Open architecture is something which many customers would like to get their hands on just because it allows them to work agnostically across the different vendors,” said the executive, who asked to not be identified because he was not authorized to speak on behalf of his company without prior approval.

“If you’re not married to one architecture, there are cost implications as well. There are availability implications as well, and the dependencies implication as well. So from a purely business point of view, this is great,” he added.

Cornelis Stands Up To Nvidia With Open Alternative

The challenge for Cornelis is that Nvidia has spent more than a year promoting its proprietary NVLink interconnect technology as the de facto method for scale-up networking to competitors building rack-scale solutions with their own accelerator chips. Such chips are posing a growing threat to Nvidia’s traditional GPU business.

Several companies have signed up so far to support NVlink, including Amazon Web Services, Arm and even Qualcomm, giving Nvidia a way to remain a crucial supplier for the AI infrastructure market and opening another revenue stream.

But even before Cornelis announced its plan to make scale-up products, there has been a growing movement by major tech companies to support an NVlink alternative. This has mainly been through the UALink Consortium, a group that is developing specifications for its open, namesake Ultra Accelerator Link standard.

“We think this is the perfect time to jump into this market. There is still much to be settled,” Spelman said, adding that the company also plans to use the Active Compute Fabric for future scale-out networking products, which will support the Ultra Ethernet standard.

Members of the UALink Consortium include major tech companies such as Amazon Web Services, AMD, Apple, Google, HPE, Intel and Microsoft as well as smaller players like Cornelis, which plans to support UAL as well as another open standard called ESUN for scale-up products.

“There are so many other compute solutions, accelerators, GPUs, XPUs that need scale-up capabilities as they build out their rack-level technology and need partners that are really focused on delivering to open standards,” Spelman said.

Cornelis Says New Architecture Will Improve Utilization

Another way Cornelis hopes to stand apart from Nvidia is how its Active Compute Fabric has been designed to improve compute utilization within rack-scale platforms.

The startup said this is accomplished by integrating programmable compute into the architecture, which allows the network to “operate on data as it moves through the system, adapt to changing workloads, offload collective operations and take on new functions as AI algorithms and software evolve.”

“We’re driving AI-capable compute into your [network interface card] and into your switch, and so you can think of this as not just compute in your network, but it is a whole other pool of compute that is meant to deliver the AI workload,” Spelman said.