Lawrence Livermore's new Lynx cluster uses a custom networking fabric from a small Intel spinoff, instead of Nvidia's InfiniBand or HPE Cray's Slingshot, in a deliberate bet to keep the agency that runs the U.S. nuclear stockpile off a single vendor.
The U.S. Department of Energy is no longer content with two ways to wire up a supercomputer. This week Lawrence Livermore National Laboratory powered on Lynx, a new cluster of 952 Dell PowerEdge servers that uses a networking fabric from Cornelis Networks, an Intel spinoff, instead of Nvidia's InfiniBand or HPE Cray's Slingshot, the two fabrics that effectively dominate the top end of high-performance computing, according to The Register's coverage.
The move looks small on paper. Lynx is a modest machine: 952 nodes is a rounding error next to the million-core monsters the Department of Energy operates at Oak Ridge and Argonne. What is not small is the strategic posture. For the first time in years, the agency that runs the country's nuclear-weapons simulations has an operational, third-lane path for connecting the compute nodes inside its machines, one that does not run through Nvidia or Hewlett Packard Enterprise.
The fabric, called Omni-Path, was originally an Intel project. Intel introduced it in 2015 as its answer to the lossless networking used to glue thousands of processors into a single coherent machine, a problem Nvidia's InfiniBand had come to dominate. Intel discontinued the product line in 2019 and spun out the team as Cornelis Networks in 2020. Five years later, the spinoff re-emerged with the CN5000 family of switches and network interface cards, and Lynx is among the first production deployments of that 400 gigabit-per-second generation, per The Register.
Lynx was commissioned by the National Nuclear Security Administration, the half of the Department of Energy responsible for maintaining the U.S. nuclear stockpile through simulation and modeling. That makes the deployment a deliberate test by a buyer that cannot afford to be locked into a single networking vendor. If InfiniBand or Slingshot pricing, roadmaps, or export controls ever tighten, the national labs still need a way to plug their compute nodes into a coherent machine. Omni-Path gives them a third option to fall back on, and a vendor to bargain with on the other two.
The compute side of Lynx is the part the deployment announcement skips past. The 952 nodes run Intel's fourth-generation Xeon Scalable processors, codenamed Sapphire Rapids, which reached the market in January 2023. In a high-performance computing world that has largely moved to fifth-generation Xeon, AMD Epyc, or Nvidia's Grace Hopper, Sapphire Rapids is now considered aging silicon. Omni-Path may be a fresh 400-gigabit fabric; the CPUs sitting behind it are not leading edge. The strategic bet is on the network, not on the compute.
Cornelis Networks itself is part of the risk calculation. The company is small, focused almost entirely on this single product line, and competes in a market where Nvidia and HPE Cray each have orders of magnitude more engineering weight behind their fabrics. Its survival depends on a handful of national-lab and research-cluster customers converting into a longer roadmap. Today's Lynx win is real, but it is one cluster, not a movement. If Cornelis falters, the Department of Energy is back to two options, with no real third lane to point to.
What to watch next is whether other Department of Energy sites follow. The agency's top-end machines, Frontier at Oak Ridge, El Capitan at Lawrence Livermore, and Aurora at Argonne, all use Slingshot or InfiniBand and will not change fabrics overnight. The interesting question is whether Cornelis can pick up a chunk of the mid-tier and capacity-cluster procurement the national labs issue every year, the kind of system where Sapphire Rapids-class iron is still the norm and the buyer is more sensitive to vendor leverage than to peak performance. A second or third deployment, on different generations of compute and at a different lab, would be the clearest signal that the third lane is actually a lane and not a one-off experiment.