Cornelis Networks is looking at accelerating AI computing, with it working on a reference architecture while also deploying its networking tech to support scientific and engineering workloads.

A newly struck partnership with chip startup NextSilicon will see the pair jointly create AI and high-performance reference architectures that will leverage Cornelis’ CN5000 fabric and its partner’s Maverick-2 compute platform.

The pair echo growing industry claims that standard Ethernet is not suited for the latency-sensitive bursts which AI inference and high-performance computing (HPC) simulation workloads generate. To address the potential congestion calamity, their reference architectures use a fabric that keeps data moving and an accelerator that keeps compute busy.

Cornelis’ CN5000 is already used in several powerful systems, most notably for the scale-out networking needs of Lynx, a supercomputer at Lawrence Livermore National Laboratory (LLNL) which was formally deployed last week. The networking vendor’s tech is also being used as part of the wider Genesis Mission, the Trump administration effort to digitize swathes of historical scientific data currently bound to tape systems for use in AI models.

NextSilicon’s Maverick-2 Intelligent Compute Accelerator (ICA) chip, meanwhile, was unveiled in October 2024 and is built on 5nm process technology from Taiwan Semiconductor Manufacturing Company (TSMC). Billed as a “novel and original computing architecture,” the startup’s hardware already counts several HPC-centric customers, including the U.S. Department of Energy, such as via the Spectra supercomputer at Sandia National Laboratories.

Brought together, the pair contend their converged stack will beat infrastructure bottlenecks that would previously have left such expensive systems underutilized.

The first phase of combined effort will see the pair validate how fabric and compute perform together across configurations to enable original equipment manufacturer (OEM) partners can “start from proven combinations rather than untested parts lists.”

The project will later expand to cover higher-end network speeds using the forthcoming 800 gigabit per second (Gb/s) CN6000 fabric.

“Operators keep telling us their most expensive systems sit idle, waiting on the network," said Lisa Spelman, CEO of Cornelis. "We built the CN5000 to end that wait. NextSilicon challenges the same kind of assumption on the compute side, so this collaboration is a natural fit. Together we can show partners and customers what a congestion-free fabric and a workload-driven compute architecture deliver as one design.”

“For decades, software had to bend to fit the processor,” added Elad Raz, founder and CEO of NextSilicon. “Maverick-2 makes the processor adapt to the software. Cornelis takes the same approach to the network. Evaluating our architectures together is the first step toward giving customers and OEM partners a faster, more efficient foundation for AI and HPC.”

Cornelis supports Stampede3

To further validate its CN5000 offering, Cornelis revealed its networking tech was deployed as part of an upgrade for the Stampede3 supercomputer.

Housed at the Texas Advanced Computing Center (TACC) in Austin, the Dell and Intel-based supercomputer is a 10-petaflop system used for simulation workloads.

Cornelis’s CN5000 deployment covers more than 600 compute nodes, with its networking tech providing improved performance and scalability to enable simulation workloads to be completed faster across weather forecasting, engineering, and data-intensive analysis.

“TACC has long been one of the most respected research computing organizations in the world, and we’re proud to continue supporting its mission,” Spelman said. “The acceptance of this upgrade marks another milestone in our collaboration with TACC and reflects our shared commitment to helping researchers tackle increasingly complex scientific challenges.”