Cornelis Omni-Path revives Intel tech for DoE supercomputers

The U.S. Department of Energy (DoE) has deployed a new supercomputer that relies on networking technology with roots in Intel’s abandoned high-performance computing (HPC) efforts. The Lynx system at Lawrence Livermore National Laboratory (LLNL) uses Cornelis Networks’ Omni-Path interconnect, offering a rare alternative to Nvidia’s InfiniBand in large-scale government computing clusters.

Lynx, built by Dell Technologies, consists of 952 nodes powered by Intel’s fourth-generation Xeon Scalable processors. While modest by DoE standards, the system stands out for its networking choice. Most DoE supercomputers use either HPE Cray’s Slingshot or Nvidia’s InfiniBand, but Lynx instead employs Cornelis’ CN5000-series Omni-Path switches and network interface cards (NICs). The deployment marks a validation of the technology, which Intel originally developed in 2015 before discontinuing it in 2019 and spinning off the division in 2020.

What happened

Cornelis Networks, the Intel spinoff behind Omni-Path, has reintroduced the technology with its CN5000 portfolio, promising 400 Gbps connectivity and near-linear performance scaling. The DoE selected the interconnect for Lynx last summer, positioning it as a high-speed alternative for non-Cray systems. The agency’s existing Cray clusters typically operate at 200 Gbps, while InfiniBand, though capable of higher speeds, remains in high demand for AI workloads.

Matt Leininger, a senior principal HPC strategist at LLNL, described the collaboration as a result of shared investment in advancing high-performance computing. The system will support classified modeling, simulation, and analysis for the National Nuclear Security Administration (NNSA). Cornelis CEO Lisa Spelman told The Register that the deployment demonstrated the CN5000 portfolio’s scaling efficiency, achieving 91% network scaling in the cluster. She suggested Lynx could outperform similarly sized systems with newer processors due to its efficient interconnect.

Why it matters

The deployment of Omni-Path in Lynx signals a potential shift in the HPC networking landscape. InfiniBand has long dominated supercomputing interconnects, but Cornelis’ technology offers a viable alternative, particularly for government and research institutions. The DoE’s adoption provides a high-profile proof point for the startup, which is already working on additional deployments, including systems with non-traditional accelerators.

Cornelis is also preparing to launch its CN6000 series later this year, targeting 800 Gbps speeds and compatibility with PCIe Gen 6.0. The new hardware will support Ethernet connectivity, broadening its appeal beyond HPC to more conventional data center environments. Unlike Nvidia, which integrates PCIe switches into its NICs to bypass bandwidth limitations, Cornelis aims to avoid added complexity and cost.

What to watch

The success of Lynx could determine whether Omni-Path gains traction beyond niche applications. Cornelis has indicated plans to scale the technology to larger clusters, with Spelman expressing confidence in its performance at 2,000 nodes and beyond. The upcoming CN6000 series will test the company’s ability to compete in both HPC and enterprise markets, particularly as PCIe Gen 6.0 adoption grows.

For now, the DoE’s endorsement provides Cornelis with a critical foothold in a market dominated by Nvidia and HPE Cray. If the technology proves reliable at scale, it could challenge InfiniBand’s dominance in supercomputing and AI infrastructure.