NVIDIA Expands Vera CPU Racks for Scientific Computing With 192-CPU Configuration
NVIDIA announced a new rack-scale platform built around its Vera CPU for high-performance scientific computing. The announcement came on Nov. 16 at the SC26 supercomputing conference. The design links 192 Vera processors with InfiniBand networking, a high-speed connection technology used to tie computers together in large clusters. Racks from Dell Technologies, HPE, Lenovo and Supermicro are expected to be available starting in early 2027.
Key details of the new rack
- The rack holds 192 Vera CPUs connected by Quantum-X800 InfiniBand networking.
- NVIDIA says the liquid-cooled rack contains about 300 terabytes of memory and 1 petabyte of storage.
- Each compute tray holds eight CPUs.
- Partner racks from Dell Technologies, HPE, Lenovo and Supermicro are expected from early 2027.
- The platform targets scientific simulations that run without GPU acceleration.
What NVIDIA says about Vera
Vera is not a new chip. NVIDIA first unveiled it in May for AI agents, reinforcement learning and data processing, including standalone CPU racks. NVIDIA calls Vera its first fully custom data center CPU. An earlier AI-focused Vera rack uses Spectrum-X Ethernet, while the new scientific-computing announcement centers on InfiniBand and software for simulation workloads.
According to NVIDIA's architecture whitepaper, each Vera processor has 88 NVIDIA-designed Olympus cores, 176 threads, and uses the Arm v9.2 instruction set. Arm is a different chip design family from the x86 processors that most data centers use. Each processor supports up to 1.5 terabytes of LPDDR5X memory and up to 1.2 terabytes per second of memory bandwidth. The announcement also mentions serviceable SOCAMM2 memory modules and reuse of NVL72 rack chassis, power distribution and cooling infrastructure.
For buyers who use x86 systems, software compatibility may matter. NVIDIA's Arm migration guidance says applications without Arm support may need recompilation, and x86-specific vector instructions can require conversion. NVIDIA says its HPC SDK 26.11, a software toolkit with compilers and libraries optimized for Vera, is available now.
How it differs from other NVIDIA products
The CPU rack is separate from NVIDIA's Vera Rubin accelerated systems. Vera Rubin NVL72 combines 36 Vera CPUs with 72 Rubin GPUs, and NVL4 pairs two Vera CPUs with four Rubin GPUs. Those systems use CPUs alongside GPU accelerators rather than focusing the rack on CPU computation.
NVIDIA also announced a Rubin HPC GPU on Nov. 16 at SC26. The company says it will target scientific simulation and analytics with sparse-data acceleration and seventh-generation NVLink connections between GPUs. Rubin HPC is expected in 2028, later than the Vera CPU racks.
What the performance tests show
NVIDIA has published specifications and selected benchmarks, but these are workload-specific vendor comparisons. In its HPC performance brief, NVIDIA reports 1.9 times the performance and 2.4 times the performance per watt on OpenFOAM's motorBike Large airflow simulation compared with a dual-socket AMD EPYC 9755 system. The comparison used dual-socket Vera hardware with LPDDR5X memory against AMD hardware with DDR5 memory.
The same brief shows a weaker result on another workload: Vera reached 0.9 times AMD's performance on the BWA genomics workload, though with 1.2 times the performance per watt. NVIDIA separately labels its weather-model performance gains as projected rather than measured. The cluster tests covered up to 12 Vera nodes and compared them with NVIDIA's earlier Grace platform. NVIDIA says competing clusters were not available for that evaluation. The launch release does not disclose rack pricing.
A confirmed customer commitment
One customer has independently confirmed a Vera commitment. The German Climate Computing Center, known as DKRZ, said on Oct. 26 that its Bull-built HLRE-5 supercomputer would include a Vera CPU partition for workloads that do not run on GPUs, alongside GB200 Grace Blackwell systems. That system specifies Quantum-2 networking rather than the new rack's Quantum-X800, and it is expected to enter service in early 2028.
What is confirmed and what is not
Confirmed information comes from NVIDIA's own announcements and DKRZ's statement. These include the rack's specifications, the partner lineup, the availability window of early 2027, and DKRZ's plan to use a Vera partition in HLRE-5.
Several points remain open. The performance numbers are NVIDIA's own vendor comparisons and cover specific workloads, so they may not reflect all real-world uses. The weather-model results are projections, not measured results. Rack pricing has not been disclosed.