Details
- NVIDIA announced the Vera Rubin platform as a rack-scale supercomputing system for science, combining native FP64 performance, CUDA-X libraries and the full NVIDIA AI stack to accelerate AI, simulation and data-intensive research workloads.
- The platform integrates Rubin GPUs and Vera CPUs connected via NVLink-C2C, ConnectX-9 SuperNICs and BlueField-4 DPUs in a direct liquid-cooled architecture, supporting high-precision simulation and AI for surrogate models, scientific foundation models and real-time analytics.
- A Vera Rubin system can scale to racks with up to 144 GPUs, delivering more than 7 exaflops of AI performance for science and 5 petaflops of native FP64, enabling larger models, higher fidelity and faster time-to-discovery for workloads such as climate modeling, CFD, quantum chemistry and energy exploration.
- Leibniz Supercomputing Centre (Blue Lion, online in 2027), NERSC (Doudna) and Los Alamos National Laboratory (Mission, Vision, Veritas) are building next-generation supercomputers on Vera Rubin to support open science, energy research, earth sciences and national security, with configurations tailored for foundation models, agentic AI and complex multiphysics simulations.
- Global system manufacturers including Bull, Dell Technologies, GIGABYTE, HPE and Supermicro will ship direct liquid-cooled Vera Rubin NVL4-based racks for AI and HPC, with initial systems expected to be available in Q4 2026; NVIDIA positions Vera Rubin as a common accelerated computing platform from single racks to large-scale centers, comparable to TOP500-class systems.
Impact
By fusing FP64 HPC with agentic AI on a single rack-scale platform, Vera Rubin reinforces the trend toward converged AI/HPC architectures in climate, energy, and national security computing. Over the next 12–24 months, deployments at leading labs are likely to influence procurement standards and push other vendors toward integrated, liquid-cooled, rack-scale systems optimized for both foundation models and traditional simulations.
