"Every AMD EPYC generation puts more cores in a node, and every one of those cores needs its own path to the network. When the network can't keep up, you pay for compute that sits waiting." Cornelis CEO Lisa Spelman's line names the most overlooked line item in HPC procurement: CPUs keep getting faster, but the network that feeds data to those cores is what decides whole-system efficiency. On 11 November 2026, Cornelis and AMD launched a joint HPC / AI-for-science program to design, validate and deliver EPYC CPUs + Instinct GPUs + Cornelis networking as a package, sold through the OEMs and integrators customers already use. The anchor is a production machine: GWDG's EMMY P4 supercomputer in Göttingen, Germany, running roughly 300 dual-socket EPYC 9745 nodes on a Cornelis CN5000 Omni-Path fabric.
CN5000 and Omni-Path: why "a path per core" matters

CN5000 is Cornelis's Omni-Path switching/adapter platform, qualified and in production on AMD EPYC 9005, validated across more than 25 platform configurations and deployed at university, research, government and industrial sites in North America and Europe; qualification now extends to AMD Instinct MI350X GPUs. It targets double-precision workloads where compute and network must scale together: CFD, climate/weather, life sciences, physics, materials science. Omni-Path shares InfiniBand's goal — low latency, controlled congestion, a dedicated communication path per core so cores never wait on the network — from a different lineage. The follow-on CN6000 is an 800G adapter with a PCIe 6.0 x16 host interface, being validated with EPYC 9006 and next-gen Instinct; Cornelis says it can give every core in a 512-core dual-socket server a dedicated network connection, citing 1.6 billion messages/second and 23% lower MPI latency than NVIDIA XDR InfiniBand (internal measurements/projections, not independently audited).
GWDG EMMY P4: production-grade proof at 300 dual-socket nodes

EMMY P4 serves NHR@Göttingen, one of nine centers in Germany's National High Performance Computing Alliance, supporting thousands of research codes across life sciences, physics, quantum computing, AI and the digital humanities, in an environment that runs ~3.5 million HPC jobs a year. Prof. Dr. Julian Kunkel, deputy head of HPC at GWDG, says EPYC plus the Cornelis fabric delivers the compute density, low latency and consistent performance the center needs. ~300 dual-socket nodes means ~600 EPYC 9745 sockets (calculated). This is a production system, not a lab demo — for buyers, "in production" carries more weight than any benchmark.
Scaling results: impressive, but read them with the caveat
Cornelis-published application results: on dual EPYC 9755 + CN5000, Siemens Simcenter STAR-CCM+ ran 8.99x faster on eight nodes than one (nine models, 112% parallel efficiency); across 132 CONVERGE CFD workloads on Dell PowerEdge R7725, CN5000 delivered 20% higher average throughput than NVIDIA NDR InfiniBand at four nodes; Cornelis cites up to 39% higher throughput across tested applications. These are Cornelis's own results; StorageReview explicitly did not run or audit them. Reading: the direction is credible (the network is indeed the bottleneck), but absolute gains depend on your codes and topology — acceptance must re-test on your own workloads.
The route war: InfiniBand vs Ethernet vs Omni-Path
HPC networking has long been InfiniBand (NVIDIA/Mellanox) versus high-performance Ethernet (RoCE); Cornelis's Omni-Path is a third route, descended from Intel's Omni-Path and carried on by Cornelis, betting on per-core dedicated paths and congestion control. If CN6000's claimed 23% latency edge over XDR InfiniBand holds, it is attractive for latency-sensitive tightly-coupled MPI codes; but ecosystem, switching and ops-tooling maturity remain InfiniBand's moat. Procurement should weigh the whole fabric's operations and code compatibility, not a single benchmark.
Procurement lessons for university / research clusters
AMD and Cornelis combine engineering, field and partner teams from architecture and benchmarking through acceptance and production; customers can run their codes at the AMD HPC Center of Excellence in Munich before committing, and the two plan joint reference designs plus ISV application performance data, extending the program to each new EPYC/Instinct generation. OEM systems pairing CN6000 with EPYC 9006 are expected from Q1 2027; Cornelis will show EPYC platforms at SC26 in Chicago, booth 1419. For Pearl River Delta universities, research institutes and CFD/weather/life-science teams: write "network and compute accepted together" into the tender, validate your MPI workloads via the CoE or a vendor POC first, then choose InfiniBand / Omni-Path / RoCE — don't let bought cores sit waiting on the network.
Sources & verification notes (accessed 2026-11-15)
- StorageReview (Harold Fritts, 2026-11-11): https://www.storagereview.com/news/cornelis-cn5000-amd-epyc-gwdg-emmy-p4-joint-hpc-program (fetched and verified directly)
- Cornelis × AMD program page: https://www.cornelis.com/partners/amd
- AMD EPYC 9005 product page: https://www.amd.com/en/products/processors/server/epyc/9005-series.html
- Calculated: ~600 EPYC sockets (300 dual-socket). Vendor claims / unaudited: 8.99x / 112%, 20% / 39% throughput, 1.6B messages/s, 23% latency, per-core path on 512-core, MI430X 288 TFLOPS FP64 (peak theoretical). CN6000 + EPYC 9006 OEM Q1 2027 and SC26 booth 1419 are vendor plans.
