Nvidia has announced the Vera CPU rack-scale platform, a new high-performance computing system built around the company’s first fully custom datacenter CPU. Vera targets demanding scientific workloads such as climate modeling, molecular dynamics and large-scale simulations that still run primarily on CPUs.
At rack scale, the platform integrates 192 Vera CPUs linked by Nvidia Quantum‑X800 InfiniBand. The liquid‑cooled Nvidia MGX design delivers more than 200 TB/s of aggregate peak memory bandwidth, aimed at keeping huge datasets moving for both traditional simulations and AI‑assisted workflows.
Each Vera CPU includes 88 custom Olympus cores, 176 threads and SOCAMM2 LPDDR5X memory, providing up to 1.2 TB/s of memory bandwidth per socket. A full rack combines around 300 TB of RAM and 1 PB of NVMe storage, with each 1U tray housing eight Vera CPUs plus storage and management.

Chassis of the Nvidia Vera CPU rack-scale platform — Image: NVIDIA
The platform reuses NVL72 rack components — chassis, power distribution and cooling — to simplify deployment and service. Quantum‑X800 InfiniBand provides the fabric, while Nvidia SHARP In‑Network Computing reduces communication overhead so CPUs spend more time on actual computation.
In Nvidia’s internal benchmarks, Vera achieved 1.9× the performance and 2.4× the performance per watt of x86 systems on the OpenFOAM motorBike Large airflow simulation. For the open-source Weather Research and Forecasting (WRF) model, Nvidia projects 1.6× speedup and 2× energy efficiency versus x86.
The German Climate Computing Center DKRZ has chosen Vera CPUs for its upcoming HLRE‑5 supercomputer, alongside GB200 Grace Blackwell systems and Quantum‑2 InfiniBand. Built by Bull and expected in early 2028, HLRE‑5 is designed for global kilometer‑scale climate simulations with lower power consumption than the current machine.
To support developers, Nvidia HPC SDK 26.11 is available now with Vera‑optimized compilers and libraries. Nvidia is also launching a public beta of its next‑generation Fortran compiler, enabling existing Fortran‑based scientific codes to run faster on Nvidia platforms without being rewritten. Vera CPU racks are expected from Dell Technologies, HPE, Lenovo and Supermicro starting in early 2027.
Source: NVIDIA, Electronics Weekly · Cover image: NVIDIA










