Nvidia announced a Vera CPU rack-scale platform in a press release on 16 November 2026, extending its first fully custom data centre CPU into systems for scientific computing. The liquid-cooled rack connects 192 Vera CPUs through Nvidia Quantum-X800 InfiniBand networking, a combination the company is targeting at work such as climate modelling and molecular simulation, where large amounts of data must move between processors.
Key points
- Nvidia specifies 192 Vera CPUs, approximately 300TB of memory and 1 petabyte of storage per rack.
- In Nvidia’s OpenFOAM test, Vera delivered 1.9x the performance and 2.4x the performance per watt of x86 systems.
- The German Climate Computing Center has selected Vera CPUs for its HLRE-5 climate supercomputer.
- Four suppliers are expected to offer Vera CPU racks beginning in early 2027.
Nvidia puts 192 Vera CPUs in one rack
Each Vera CPU has 88 custom Olympus cores and 176 threads, according to Nvidia, which rates its LPDDR5X memory at up to 1.2TB/s of bandwidth. The company specifies approximately 300TB of memory and 1 petabyte of storage for the complete rack. Eight CPUs fit into each 1U tray alongside NVMe storage and management components, and liquid cooling runs through the rack.
Nvidia rates the rack’s aggregate peak memory bandwidth at more than 200TB/s. Bandwidth is the rate at which data can pass through memory, much as the capacity of a road depends on how much traffic can pass a point in a given time. For a simulation repeatedly reading large datasets, that rate is a separate constraint from how many calculations its CPUs can perform. Nvidia also says the design reuses chassis, power distribution and cooling components from its NVL72 racks.
The CPUs communicate over Quantum-X800 InfiniBand. Nvidia says its SHARP in-network computing reduces the time processors spend coordinating during complex simulations, moving some of that work into the network rather than leaving every exchange to the CPUs. That connection is part of the rack’s design: adding processors to a simulation also increases the need to pass information between them.
AI Affairs previously reported Supermicro’s liquid-cooled rack with 256 Vera CPUs. Nvidia is presenting the newly announced 192-CPU configuration for high-performance computing, with the memory, network and scientific software specified together.
OpenFOAM tests and Nvidia’s weather projection
In its own tests using OpenFOAM’s motorBike Large airflow simulation, Nvidia reported 1.9x the performance and 2.4x the performance per watt for Vera against x86 systems. OpenFOAM runs fluid-flow simulations, so this test exercises the platform on a defined airflow problem rather than on every kind of scientific calculation. The performance-per-watt figure measures work against power consumption in that test.
Nvidia also benchmarked the open-source Weather Research and Forecasting model on Vera CPUs and projected performance gains of 1.6x and energy-efficiency gains of 2x compared with x86. Those weather-model figures are the company’s projections, whereas the OpenFOAM figures are results it reported from its tests. Both comparisons concern the named workloads and the x86 systems Nvidia used as its reference.
Checking the weather before leaving home could benefit from faster simulations if Nvidia’s projected gains on the weather model carry through to the calculations behind forecasts.
For software builders, Nvidia says HPC SDK 26.11 is available with compilers and libraries optimised for Vera. The company has also announced a public beta of a next-generation Fortran compiler developed with high-performance computing centres and national laboratories. Fortran remains a language used in scientific computing, and Nvidia says the compiler is intended to speed existing programs on its platforms without requiring their code to be rewritten.
DKRZ selects Vera for HLRE-5
DKRZ, the German Climate Computing Center, has selected Vera CPUs for its HLRE-5 supercomputer, Nvidia said. The planned machine will also include Nvidia GB200 Grace Blackwell systems and Quantum-2 InfiniBand networking. Nvidia says HLRE-5 will support climate simulation and AI research, including global climate simulations at kilometre scale, with lower energy consumption than DKRZ’s current system.
The DKRZ selection places Vera in a planned scientific installation that combines CPUs with other computing systems, rather than treating the 192-CPU rack as the whole machine. Its Quantum-2 network is also distinct from the Quantum-X800 networking Nvidia specifies for the newly announced Vera rack.
Dell Technologies, HPE, Lenovo and Supermicro are expected to offer Vera CPU racks beginning in early 2027, according to Nvidia. That timetable would give organisations building scientific systems several suppliers for the rack configuration, alongside the compiler and library package Nvidia says is already available.
Bull is building DKRZ’s HLRE-5 supercomputer, which Nvidia expects to be available in early 2028.