Key points
- Nvidia brings Vera CPUs to scientific computing
- Nvidia's x86 comparison rests on its own tests
- Germany's climate center picks Nvidia's Vera
Nvidia (NVDA) is bringing its Vera CPU to scientific computing in a liquid-cooled rack that holds 192 of the processors. The company expects the racks to be available from Dell Technologies (DELL), HPE (HPE), Lenovo, and Supermicro (SMCI) beginning in early 2027.
Nvidia announced the rack on Nov. 16 at the SC26 supercomputing conference in Chicago. It also compared Vera's performance with x86 server processors, using its own tests and projections. Intel (INTC) and AMD (AMD) make server processors based on the x86 design.
A CPU is the general-purpose processor that runs a computer's operating system and much of its software. Nvidia built its business on GPUs, which handle the parallel math behind AI models. The company calls Vera its first fully custom data center CPU.
Vera runs on 88 cores Nvidia designed itself
Each Vera chip has 88 cores, which Nvidia calls Olympus, and can run 176 threads at once. A core is a processing unit inside the chip, and a thread is a stream of instructions it works through. Nvidia says Vera has "full Arm compatibility," so it can run software written for chips based on Arm Holdings' (ARM) architecture.
Nvidia's earlier Grace data center CPU uses Arm's Neoverse V2 cores, a design Arm licenses to chipmakers. Vera's Olympus cores are Nvidia's own design.
Vera can work alongside GPUs or on its own. Nvidia's Vera Rubin NVL72 system combines 36 Vera CPUs and 72 Rubin GPUs in one rack. Vera is also available in single- and dual-socket servers and full racks.
AI labs got the first Vera systems
The first Vera systems went to Oracle Cloud Infrastructure and three AI labs: Anthropic, OpenAI, and SpaceXAI, Nvidia said in a blog post updated on Aug. 27. AWS had also received its first Vera server by then.
Nvidia has pitched Vera as its first CPU built for AI agents, software that carries out tasks on its own. SpaceXAI, part of SpaceX (SPCX), is adopting the chip to run the orchestration and code execution behind its Grok models.
OpenAI runs agent sandbox workloads on Vera CPUs in Oracle's (ORCL) cloud, where Vera became generally available in October. Meta Platforms (META) plans to expand its use of Vera CPUs as part of a partnership it announced with Nvidia in October.
The new rack is built for scientific simulations
The rack announced Nov. 16 holds 192 Vera CPUs with about 300TB of memory and 1 petabyte of storage. Nvidia's Quantum-X800 InfiniBand networking links the chips. Nvidia puts the rack's aggregate peak memory bandwidth at more than 200TB/s.
Each 1U tray holds eight Vera CPUs plus storage. The rack reuses the chassis, power distribution, and cooling from Nvidia's NVL72 design.
"Scientific simulations need CPUs that can keep large datasets moving and sustain performance as applications scale," said Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, in the release.
Nvidia's x86 comparison comes from its own tests
In an airflow simulation run on OpenFOAM, an open-source fluid dynamics program, Vera delivered 1.9 times the performance and 2.4 times the performance per watt of x86 systems, Nvidia said. The release doesn't name the x86 chips. The only x86 system listed in Nvidia's Vera performance brief is a two-socket server with AMD's EPYC 9755, a Turin-generation chip. The brief compares Vera with what it calls the "latest competition."
For weather forecasting with the WRF model, Nvidia projects 1.6 times the performance and twice the energy efficiency of x86. Nvidia describes those weather gains as projected.
Germany's climate center picked Vera for 2028
DKRZ, the German Climate Computing Center, has selected Vera CPUs for its HLRE-5 supercomputer. The system will also use Nvidia's GB200 Grace Blackwell systems and Quantum-2 InfiniBand networking. Bull is building it, and it's expected to be available in early 2028.
Nvidia is also releasing a Fortran compiler
Much scientific software is written in Fortran, a programming language that dates to the 1950s. A compiler translates that code into instructions a chip can run.
Nvidia is introducing a public beta of a next-generation Fortran compiler, built with high-performance computing centers and national labs. Nvidia says it lets developers run existing Fortran programs faster on its platforms "without needing to rewrite their code." Its HPC SDK 26.11, a set of compilers and libraries optimized for Vera, is available now.
What isn't known yet
The release doesn't give a price for the rack or say how many racks the four server makers plan to build. DKRZ has selected Vera CPUs, though the release doesn't say whether its system will use the new rack.












