Scale AI and HPC faster with AMD + Cornelis
Design, validate and deploy AMD EPYC™ and AMD Instinct™ systems with Cornelis networking, through the OEMs and system integrators you already use.
More choice in compute, validated on your applications
Our collaboration with AMD helps HPC and AI for science customers design, validate and deploy systems that pair AMD EPYC™ CPUs and AMD Instinct™ GPUs with Cornelis® networking. Customers get more choice in compute, systems validated on their own applications, and a direct path to buy and deploy through the OEMs and system integrators they already use.
Qualified and in production
Cornelis CN5000 is qualified and in production on AMD EPYC™ 9005 Series platforms, validated on more than 25 platform configurations, and deployed at university, research, government and industrial sites across North America and Europe. Cornelis has also extended CN5000 qualification to AMD Instinct™ MI350X Series GPUs.
Built for scale
Our solutions target scientific and engineering workloads where compute and networking must scale together: computational fluid dynamics, climate and weather, life sciences, physics, materials science and other FP64-intensive codes.
Expanding customer choice
AMD and Cornelis are working together to expand customer choice through open, interoperable AI and HPC solutions. Our collaboration gives organizations greater flexibility to select the compute and networking technologies that best meet their performance, scalability, and infrastructure needs.
Demonstrated performance across HPC workloads
AMD EPYC™ processors with Cornelis CN5000 have demonstrated strong performance and scaling across scientific and engineering applications including Ansys Fluent, OpenFOAM, Siemens Simcenter STAR-CCM+, CONVERGE CFD and WRF.
Ansys Fluent®
115%
average parallel efficiency at 8 nodes across four Fluent models, peaking at 129%, on dual AMD EPYC™ 9755 servers with CN5000.
Siemens Simcenter™ STAR-CCM+™
112%
geometric-mean parallel efficiency at 8 nodes across nine STAR-CCM+ models, an 8.99× speedup, on dual AMD EPYC™ 9755 servers with CN5000.
CONVERGE CFD
~90%
scaling efficiency at 2 nodes, and about 70% at 4, across 132 workloads in 12 industry verticals on Dell R7725 servers with dual AMD EPYC™ 9755 and CN5000.
Deployed at GWDG
"AMD EPYC processors and the Cornelis fabric give us the compute density, low latency and consistent performance to serve them well, so our users can focus on their research."
Support from design to production
AMD and Cornelis technical teams work together on your deployment, from architecture guidance through production operations.
How Engagement Works
Solutions ship through OEM platforms. Joint technical engagement covers:
Architecture guidance and pre-deployment planning
Workload testing in Munich
Escalation support during and after deployment
The latest from AMD and Cornelis
View All
OpenFOAM® Performance and Scalability with Cornelis Networks® OPX on AMD EPYC™ 7003-Series Processors
Beyond the Benchmark: How OEM-Validated AI and HPC Fabrics Reduce Deployment Risk
What OEM validation adds on top of a benchmark sheet, and why it matters before a fabric reaches your data centre.
What We Heard at Advancing AI 2026, and What We're Building With AMD
This week our team joined AMD in San Francisco for Advancing AI 2026.
Frequently asked questions
-
Does Cornelis networking work with AMD ROCm™?
Yes. Cornelis has published GPU-to-GPU tests on CN5000 running ROCm™ 7.2.3 with Open MPI and Cornelis OPX software, moving data directly between AMD Instinct™ MI300X GPU buffers. The CN6000 family ships with a libfabric-based software stack designed to be compatible with AMD ROCm. -
Which AMD GPUs has Cornelis networking been tested with?
Cornelis has published results for AMD Instinct™ MI300X on CN5000: up to 94% of line rate unidirectional and 93% bidirectional between GPU buffers, across two nodes connected through a CN5000 switch. The CN6000 reference architecture is built for the AMD Instinct™ MI400 Series. -
Do my MPI and RCCL applications need changes to run on Cornelis?
No. Open MPI, MPICH, MVAPICH, NCCL, and RCCL run on the Cornelis OPX software stack without modification. OPX is open source and built on libfabric, the OpenFabrics interface. -
Can I evaluate AMD + Cornelis before I buy?
Yes. Qualified teams can test workloads at the Munich Center of Excellence or request a guided evaluation.