On November 16, 2026, NVIDIA announced a liquid-cooled rack for scientific computing that links 192 Vera CPUs with its high-speed Quantum-X800 InfiniBand network. The company plans to offer it through Dell Technologies, HPE, Lenovo, and Supermicro starting in early 2027. It is aimed at workloads such as climate modeling and computational fluid dynamics, where many CPUs process large volumes of data in parallel.
The new rack is designed to raise overall system performance by combining fast memory supply to the CPUs with efficient communication between the nodes that share the computation. NVIDIA cites a 1.9x improvement in fluid dynamics and 1.6x in weather modeling. However, the former is a result from NVIDIA's own testing, while the latter is a projection, so the two figures rest on different evaluation methods. Below, we look at the hardware configuration and the comparison conditions behind each number, based on the published performance materials.
High-speed memory and inter-node networking tie 192 Vera CPUs together
Each Vera CPU has 88 custom-designed "Olympus" cores and can handle 176 threads. For memory, it uses LPDDR5X on replaceable SOCAMM2 modules, with up to 1.2TB/s of memory bandwidth. Memory bandwidth is the amount of data that can be transferred between the CPU and memory per unit of time. Scientific computing often involves repeatedly reading large amounts of data, so how quickly data can be fed to the compute cores largely determines performance.
According to NVIDIA's announcement, the new rack uses a liquid-cooled MGX design and holds about 300TB of memory and 1PB of storage. Total memory bandwidth across the rack exceeds 200TB/s. Note that the roughly 300TB figure refers to the capacity available for holding data, whereas the 200TB/s figure is the sum of each CPU's peak memory bandwidth.
Another key element is the network linking the CPU nodes. Nodes within the rack are connected via Quantum-X800 InfiniBand, which also supports in-network computing through SHARP. SHARP executes operations such as the aggregations required in parallel processing on the network side, reducing the time CPUs spend on communication and synchronization. The aim is to speed up not only data transfer between CPU and memory, but also large-scale simulations in which nodes exchange data frequently.
When it comes to the number of CPUs per rack, the figure in the announcement should be distinguished from the maximum design configuration. The rack announced on November 16 has 192 CPUs, but NVIDIA's performance materials (page 18) show a configuration that scales up to 256. A 1U tray holds four two-socket nodes, for a total of eight Vera CPUs. Housing 32 trays would yield 256 CPUs. Each tray also integrates NVMe storage and management functions.
The chassis, power distribution, and cooling infrastructure reportedly share components with the existing NVL72 rack, which is meant to simplify deployment and maintenance.
NVIDIA also evaluated performance with multiple nodes working together. On page 22 of the materials, three representative high-performance computing workloads were run on up to 12 Vera nodes, showing up to 2.6x the cluster performance of the previous-generation platform. Because this evaluation includes data exchange and synchronization between nodes, its conditions differ from a single-node performance comparison.
1.9x in fluid dynamics, 1.6x in weather modeling
NVIDIA's concrete examples of performance gains are the computational fluid dynamics software OpenFOAM and the weather forecasting model WRF.
OpenFOAM's "motorBike Large" is a benchmark that analyzes airflow around a motorcycle on a large computational mesh. The calculation requires repeatedly solving sparse matrices, which generates frequent memory data supply and communication between CPU cores. NVIDIA's performance materials also describe it as a workload in which memory bandwidth, cache efficiency, and parallel scalability are important.
According to NVIDIA's announcement and page 23 of the performance materials, the comparison conditions and results are as follows.
| Use case / software | Vera configuration | Comparison target | Performance | Performance per watt | Evaluation method |
|---|---|---|---|---|---|
| OpenFOAM v2606 / motorBike Large | Vera C2, 2 sockets | AMD EPYC 9755, 2 sockets | 1.9x | 2.4x | NVIDIA test results |
| WRF 4.8.0 (weather forecasting model) | Vera C2, 2 sockets | AMD EPYC 9755, 2 sockets | 1.6x | 2.0x | NVIDIA projection |
Both used GCC 16.2 as the compiler, and the AMD EPYC comparison system was a Supermicro H14DSG-OD.
The comparison evaluates not only Vera's processing performance but also performance relative to power consumption. OpenFOAM achieved 1.9x performance and 2.4x performance per watt, while WRF is projected at 1.6x and 2.0x, respectively. Both are comparisons between two-socket configurations and are a separate metric from the cluster performance on up to 12 nodes mentioned above.
On the other hand, different results emerged for workloads where core count directly affects performance. In its performance materials, NVIDIA compared a two-socket Vera system with 176 cores in total against a two-socket AMD EPYC system with 256 cores using the genomic sequence analysis software BWA. Because each thread in BWA can process sequence data relatively independently, the EPYC with more cores completed sequence alignment slightly faster. Across the full dataset evaluated, however, NVIDIA says Vera achieved 1.2x the performance per watt.
These results also show that Vera's advantage is not uniform across all workloads but depends on their characteristics. The factors that determine performance differ between processing that is sensitive to memory bandwidth and processing in which many cores can run independent tasks in parallel.
Germany's DKRZ will split work between CPUs and GPUs
One research institution adopting the Vera CPU is the German Climate Computing Center (DKRZ). Its next supercomputer, "HLRE-5," will combine a CPU partition using Vera CPUs with a GPU partition using GB200 Grace Blackwell. The GPU partition will handle the main simulations and AI computation, while the CPU partition will take on processing that is not suited to GPU execution.
According to DKRZ's contract announcement, HLRE-5 will be the first core system at the center to use the Arm architecture. All nodes will be connected by 400Gbit/s Quantum-2 InfiniBand, supporting large-scale parallel computation for global climate models at kilometer-scale resolution.
Note that the HLRE-5 announcement predates NVIDIA's latest one. DKRZ announced its contract with Bull on October 26, 2026, and the system is scheduled to go into operation in early 2028. What NVIDIA's November 16 announcement newly revealed is the configuration of the HPC rack with 192 Vera CPUs and its availability plans.
For HLRE-5 as a whole, the plan is to run two global climate simulations at 5km horizontal resolution simultaneously. DKRZ also sets a goal of cutting the energy needed to compute one simulated year by about 75% compared with the current Levante system. This is a performance and efficiency target for the whole of HLRE-5, including both the CPU and GPU partitions, not for the Vera CPU alone.
HPC SDK is available; next-generation Fortran compiler in public beta
Ahead of hardware shipments, NVIDIA is also building out the development environment for Vera. The NVIDIA HPC SDK 26.11, which includes compilers and libraries, is already available, and a public beta of a next-generation Fortran compiler has also been announced.
Fortran is a programming language long used in scientific computing for weather, fluid dynamics, materials science, and more, and research institutions have accumulated large amounts of existing Fortran code. NVIDIA says it is working with major HPC centers and national laboratories to create an environment in which such existing programs can run quickly on NVIDIA platforms without being substantially rewritten.
For research institutions, the SDK's availability and the compiler's public beta offer a chance to port and validate existing programs before the actual hardware is installed. The racks themselves are due to ship through Dell Technologies, HPE, Lenovo, and Supermicro starting in early 2027, while DKRZ's HLRE-5 aims to go live in early 2028, roughly a year later.
When adopting Vera, it will be important to understand whether existing scientific software is constrained by memory bandwidth, inter-node communication, or CPU core count. In systems that also use GPUs, how work is divided between CPUs and GPUs affects both performance and power efficiency. NVIDIA's two-socket benchmarks and cluster performance comparisons need to be evaluated with these workload-specific characteristics in mind.
