Phoronix has run the first broad independent set of HPC benchmarks on a dual-socket NVIDIA Vera server, and the 176-core Arm system held its own against 256-core EPYC and Xeon boxes, winning outright on memory-bound codes. The result is real, but it is against two-year-old x86 parts and on a vendor-tuned software stack.
What was announced
- Phoronix tested a remote dual NVIDIA Vera server with 176 cores and 352 threads in total, fitted with 16 x 96GB LPDDR5-9600 Samsung modules.
- The comparison systems were dual AMD EPYC 9575F (128 cores) and dual EPYC 9755 (256 cores), both with 24 x 64GB DDR5-6400, and dual Intel Xeon 6980P (256 cores) with 24 x 64GB MRDIMM-8800.
- The Vera server ran NVIDIA's recommended configuration: a modified Ubuntu 24.04 LTS with a patched Linux 6.17 kernel using 64K pages and GCC 15.3. The x86 servers ran stock Ubuntu 26.04 LTS with a GCC 15 based compiler.
- Vera took a large first place in OpenFOAM CFD on the large drivaerFastback mesh, and Phoronix described its WRF weather-model result as the best it has recorded. It also led in HPCG, which is driven by memory bandwidth.
- Results were more mixed on less memory-bound work: the EPYC 9755 was ahead on the small OpenFOAM mesh execution time, LAMMPS placed Vera between the EPYC 9575F and Xeon 6980P, and EPYC Turin led libxsmm at the largest GEMM size.
- Vera typically peaked at 3.2 to 3.3GHz, against 3.9GHz for the Xeon 6980P, 4.3GHz for the EPYC 9755 and 5.0GHz for the EPYC 9575F.
- Across the HPC tests the two Vera CPUs averaged 842W with a measured peak of 1,044W, a similar peak to the dual EPYC 9755. The dual Xeon 6980P drew the most power.
- Phoronix notes that EPYC Turin and Xeon Granite Rapids are two years old, that EPYC Venice results with MRDIMM-12800 support are still to come, and that Intel's Xeon Diamond Rapids is not expected until 2027.
The ETON view
The useful lesson here is not that Arm has beaten x86. It is that for a large class of HPC and simulation work, memory bandwidth per core now matters more than core count or clock speed. Vera won where the code streams data, such as CFD, weather and HPCG, with 88 cores per socket running at about 3.3GHz against parts with far more cores and much higher clocks. Where the work is more compute-bound, EPYC Turin stayed level or ahead. Anyone sizing a cluster should profile their own codes for memory bandwidth before choosing a platform, because that profile decides the answer more than the brand on the CPU.
There are three cautions before anyone redraws a roadmap on the back of this. First, the x86 comparison is with generation-old parts. AMD's Venice brings much faster MRDIMM support and is only now shipping, so the fairest test has not been run yet. Second, Vera was measured on NVIDIA's own patched kernel and tuned stack. For a lab that means validation work across schedulers, compilers, licensed solvers and monitoring tools before production, and that is a cost in staff time that does not appear in a benchmark chart. Third, Vera's memory is LPDDR5X modules rather than standard DDR5 RDIMMs, so an existing pool of spares, a familiar upgrade path and the second-hand value of the memory do not carry across in the same way.
Our view is that Vera is now a credible HPC option rather than an AI-only curiosity, but the buying decision should wait for Venice results and a pilot on your own workloads. In the meantime, with server CPU lead times already stretched, the practical move for most UK HPC and hosting buyers is to keep proven EPYC and Xeon platforms in the plan and test Arm alongside them. We can quote current EPYC and Xeon platforms across HPE, Dell, Lenovo, Supermicro and GIGABYTE side by side, help scope a pilot node for Arm, and price buyback on the cluster you are replacing.
Related infrastructure
Category: Servers · Vendor: NVIDIA, AMD, Intel · Technology: NVIDIA Vera, Arm64, LPDDR5X, EPYC 9005, Xeon 6 · Last verified: 08 Oct 2026 · ~3 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
