NVIDIA introduces Vera CPU platform for high-performance computing

The new rack-scale system features custom data center processors designed to improve efficiency and throughput for complex scientific simulations.

Image: NVIDIA Blog

NVIDIA has unveiled the Vera CPU, a custom-designed data center processor intended to accelerate high-performance computing (HPC) workloads. The new rack-scale platform integrates 192 Vera CPUs interconnected by NVIDIA Quantum-X800 InfiniBand networking, aiming to address performance bottlenecks in fields such as climate modeling, molecular dynamics, and agentic science. The system is built on an NVIDIA MGX design and utilizes liquid cooling to manage the thermal demands of its high-density configuration.

Each Vera CPU is equipped with 88 custom NVIDIA Olympus cores and 176 threads. The architecture incorporates compact, serviceable SOCAMM2 LPDDR5X memory, which provides up to 1.2TB/s of bandwidth to support data-intensive applications. According to NVIDIA, the platform delivers an aggregate peak memory bandwidth exceeding 200TB/s across the rack, which is intended to sustain performance as scientific simulations scale to larger datasets.

In internal testing, NVIDIA reported that the Vera platform achieved 1.9x the performance and 2.4x the performance per watt compared to x86-based systems when running the OpenFOAM motorBike Large airflow simulation. Similar efficiency gains were observed in benchmarks of the Weather Research and Forecasting model, where the company projected a 1.6x performance increase and a 2x improvement in energy efficiency over existing x86 alternatives.

The platform is designed for modular deployment, reusing components from the NVL72 rack architecture, including power distribution, chassis, and cooling infrastructure. Each 1U tray within the rack houses eight Vera CPUs alongside NVMe storage and management components. NVIDIA SHARP In-Network Computing is also utilized to reduce the time processors spend coordinating tasks during complex simulations.

The German Climate Computing Center (DKRZ) has selected the Vera CPU for its upcoming HLRE-5 supercomputer, which is being constructed by Bull. Expected to be available in early 2028, the system will combine Vera CPUs with NVIDIA GB200 Grace Blackwell systems and Quantum-2 InfiniBand networking to support climate and AI research. To assist developers, NVIDIA has released the HPC SDK 26.11, which includes compilers and libraries optimized for the new architecture.

Sources

  1. HPCwireNVIDIA Expands Vera CPU Platform to Accelerate Scientific Discovery
  2. NVIDIA BlogNVIDIA CUDA-X Libraries Give Agents the Tools to Accelerate Scientific Discovery