GWDG's EMMY P4 Supercomputer Pairs 300 AMD EPYC Nodes with Cornelis Networking

Germany's GWDG has deployed EMMY P4, a supercombining approximately 300 dual-socket AMD EPYC 9745 nodes connected through Cornelis CN5000 Omni-Path networking, and the installation anchors an expanded AMD–Cornelis partnership that goes well beyond selling compatible parts.

The two companies are building a joint go-to-market program covering system design, application benchmarking, validation, acceptance testing, and production deployment through existing manufacturers and system integrators. Customers can also test their own applications at AMD's HPC Center of Excellence in Munich before committing to a configuration. That last detail matters: it targets a persistent supercomputing problem where buying more processors does not guarantee proportionally faster research.

The stakes are real for research computing budgets. Supercomputing procurement has long forced buyers to evaluate processors and networks separately, on spec sheets, and hope the combination performs once real scientific code runs. AMD and Cornelis are now proposing to validate the pieces together against actual applications instead.

Why does the network matter as much as the compute?

Scientific simulations typically divide a large problem across many servers. Each server calculates part of the result, then exchanges information with the others. Computational fluid dynamics, for instance, requires neighboring regions of a simulated flow to share updated conditions as the calculation progresses.

Those exchanges make network latency, message-handling capacity, and congestion consequential. A processor waiting for another node's data contributes little useful computation, however fast it performs arithmetic. Message Passing Interface, or MPI, is the widely used software interface for coordinating this distributed work.

In this arrangement, Cornelis supplies the interconnect fabric connecting the nodes while AMD supplies the compute. Lisa Spelman, CEO of Cornelis, framed the core problem directly: each new EPYC generation increases the number of cores needing access to the network. Her point is that adding compute capacity can leave expensive resources idle if communication cannot keep pace, which makes workload-specific validation essential.

The takeaway: a supercomputer is only as fast as its slowest data path, not its fastest processor.

What does GWDG's EMMY P4 actually contain?

GWDG, the scientific computing organization in Göttingen, supports NHR@Göttingen, one of Germany's nine national high-performance computing centers. Its users run applications spanning life sciences, physics, quantum research, AI, and the humanities. Cornelis says the broader environment handles approximately 3.5 million HPC jobs annually.

AMD's EPYC 9745 specifications list 128 physical cores per processor, which puts 256 cores in each dual-socket node. Keeping that many cores productive across hundreds of nodes turns communication efficiency into a system-level concern, not a per-server one.

GWDG's installation documentation describes an initial group of 252 nodes with:

• 768 GiB of DDR5 memory per node
• 3.8 TiB of local NVMe storage per node
• CN5000 adapters with 200G links

The announcement describes the larger EMMY P4 deployment as approximately 300 nodes.

One distinction is worth making explicit. EMMY P4 demonstrates EPYC processors and Cornelis networking in a scientific computing environment; it does not mean the deployed CPU nodes contain AMD Instinct accelerators. The partnership's GPU work is separate from this CPU installation, though both are part of the same roadmap.

The takeaway: EMMY P4 is a production-scale proof that the CPU-and-network pairing works under real research load.

How does application testing move beyond compatibility?

Cornelis says CN5000 is qualified for production with AMD EPYC 9005 processors across more than 25 validated platform configurations in North America and Europe. The companies have also extended CN5000 qualification to AMD Instinct MI350X Series accelerators.

For buyers, application testing can reveal whether an architecture fits the work that actually runs in their organization. A large simulation with frequent synchronization can behave very differently from an AI workload that spends more time inside accelerator kernels. Memory capacity, software libraries, node configuration, and network behavior all influence the result.

Cornelis reports two vendor benchmarks that illustrate the approach:

• A selected set of 132 CONVERGE computational fluid dynamics workloads achieved 20% higher average throughput on four nodes than a comparison using NVIDIA NDR InfiniBand.
• A STAR-CCM+ workload running on EPYC 9755 processors showed nearly ninefold scaling from one to eight nodes.

These are vendor-reported application results from particular test configurations. The supporting Cornelis technical white papers are the relevant reference for test conditions, and the figures do not establish the same advantage for every code or cluster. The practical value of the Munich testing program is that customers can assess their own workloads against that backdrop rather than trusting either vendor's highlight reel.

The takeaway: benchmark your own code, because a 20% gain on someone else's workload may not transfer to yours.

What comes next on the hardware roadmap?

The companies are also preparing for another hardware generation. Cornelis positions its upcoming CN6000 around an 800G adapter and a PCIe 6.0 x16 interface, with OEM systems combining CN6000 and AMD EPYC 9006 processors expected in the first quarter of 2027. Those systems remain a roadmap item, separate from the deployed CN5000-based EMMY P4 infrastructure.

AMD's accelerator strategy also includes the ROCm software stack, which provides tools and libraries for GPU computing. Supporting scientific AI requires that software layer as much as the processors and network: researchers need applications, mathematical libraries, and communication software that work together.

Traditional scientific computing also leans heavily on double-precision, or FP64, arithmetic. That differs from the lower-precision calculations used in many AI workloads, so a platform's suitability for research depends on more than its headline AI performance.

The takeaway: the 2027 CN6000-plus-EPYC 9006 generation signals this partnership is a long-term architecture bet, not a one-off deployment.

What does this mean for research computing buyers?

The expanded AMD–Cornelis partnership is ultimately a bid to make complete research systems easier to evaluate and deploy. GWDG supplies a production example of the CPU-and-network combination, while joint application testing and future accelerator support broaden the potential use cases beyond Göttingen.

For research centers weighing procurement decisions in the next several years, the structure of the offer matters as much as the hardware. Joint validation, benchmarking and acceptance testing shift some of the integration risk from the buyer to the vendors. That said, the NVIDIA comparison benchmark shows Cornelis is competing directly against the incumbent InfiniBand ecosystem, and vendor-reported numbers deserve scrutiny.

The consequential measure for any research center is how quickly a validated system completes useful work — not how many processors it contains.