Nvidia said on Nov. 16 that it is expanding its Vera CPU platform with a rack-scale high-performance computing system for scientific simulation workloads, according to an announcement from the company1.

The move matters because Nvidia is positioning Vera beyond the agentic AI role it outlined earlier and into memory-intensive scientific computing, with vendor-supplied claims on bandwidth, power efficiency and support for existing Fortran-based software stacks. Nvidia's announcement includes no independent test results.

The new system combines 192 Vera CPUs with Quantum-X800 InfiniBand networking1 in a liquid-cooled rack that Nvidia said is aimed at climate modeling, molecular dynamics and other "unaccelerated" scientific workloads. The company said the rack provides more than 200TB/s of aggregate peak memory bandwidth, about 300TB of memory and 1 petabyte of storage1.

That extends a Vera program Nvidia introduced earlier in the year with an agentic AI-focused Vera launch4 that described a rack with 256 liquid-cooled Vera CPUs. Automatica reported earlier on Nov. 16 that Nvidia separately announced its Rubin HPC simulation GPU for 2028 without independent performance data3, giving Nvidia two distinct scientific-computing messages on the same day: one centered on a future GPU and one on a CPU rack available earlier.

System details and performance claims

According to the NVIDIA Newsroom release1, each Vera CPU has 88 custom Olympus cores, 176 threads and SOCAMM2 LPDDR5X memory delivering up to 1.2TB/s of bandwidth. Nvidia said each 1U tray integrates eight Vera CPUs along with NVMe storage and management, and that the rack reuses NVL72 chassis, power distribution and cooling components.

Nvidia said the rack uses Quantum-X800 InfiniBand networking and SHARP in-network computing, while the same release emphasized software support through HPC SDK 26.11, which Nvidia said is available now1, and through a public beta of a next-generation Fortran compiler.

The benchmark figures disclosed so far are all Nvidia's own. The company said that, in its tests, Vera delivered 1.9x the performance and 2.4x the performance per watt of x86 systems on OpenFOAM's motorBike Large airflow simulation1. Nvidia also said benchmarking of the Weather Research and Forecasting model showed projected gains of 1.6x in performance and 2x in energy efficiency compared with x861.

Those numbers identify Nvidia's stated baseline and workloads, but the release does not include the detailed test configurations behind them. Nvidia said a separate HPC performance brief contains benchmark results and configurations, but that document was not available in the retrieved materials for review. The announcement does not include pricing.

Availability and DKRZ selection

Nvidia said Vera CPU racks are expected to be available from Dell Technologies, HPE, Lenovo and Supermicro beginning in early 20271. The company also said DKRZ, the German Climate Computing Center, has selected Vera CPUs for its HLRE-5 supercomputer1, which Nvidia said Bull is building and is expected to be available in early 2028.

According to Nvidia, HLRE-5 will use Vera CPUs alongside GB200 Grace Blackwell systems and Quantum-2 InfiniBand networking for climate simulation and AI research. Nvidia said the system is intended to enable global kilometer-scale climate simulations with lower energy consumption than DKRZ's current system, but the release does not quantify that comparison.

The company described the announcement as an expansion of Vera for scientific discovery, with emphasis on memory bandwidth, rack integration, InfiniBand and Fortran support for existing scientific applications. Detailed benchmark configurations were not included in the retrieved announcement materials.