NVIDIA Vera CPU Powers New Los Alamos Supercomputers for Agentic Scientific AI

Los Alamos National Laboratory (LANL) is deploying three new supercomputers — Mission, Vision and Veritas — built in partnership with HPE and NVIDIA to accelerate scientific discovery through agentic artificial intelligence. The systems will leverage NVIDIA Vera CPUs alongside NVIDIA Rubin GPUs and NVIDIA Quantum-X800 InfiniBand networking within the HPE Cray Supercomputing GX5000 architecture.

Mission will incorporate 2,300 standalone NVIDIA Vera CPUs using HPE Cray GX240 blades along with NVIDIA Vera Rubin GPU nodes, while Veritas will feature approximately 1,150 standalone Vera CPUs to complement Vera Rubin nodes. Both Mission and Vision are expected to become operational in 2027. Mission will replace Crossroads for classified national security workloads as part of the National Nuclear Security Administration's Advanced Simulation and Computing program, while Vision will support fundamental science including materials science, nuclear science, energy modeling, biomedical research and artificial intelligence.

Early testing demonstrates significant performance gains from the Vera CPU architecture. LANL's benchmarks show the Vera CPU delivers 7 times higher performance on URSA (Universal Research and Scientific Agent) workloads compared to CPUs in the Crossroads supercomputer, and over 3 times better performance on Branson, an open-source Monte Carlo heat transfer simulation tool. A single Vera CPU provides more than 4 times the memory per core and 6 times the memory per node compared to single-socket x86-based CPUs.

The Vera CPU's performance gains stem from its custom Olympus core, LPDDR5 memory architecture and fast on-chip fabric. Veritas will serve the Laboratory Directed Research and Development program and test these technologies for deployment in larger systems at LANL. The systems enable researchers to develop AI agents capable of forming hypotheses, selecting tools, launching simulations, analyzing outputs and refining subsequent steps — a capability demonstrated through LANL's work on URSA, a modular, feedback-driven AI framework designed to help scientists brainstorm hypotheses, plan experiments, run simulations and analyze results.

The partnership reflects more than a decade of collaboration between LANL and NVIDIA on CPU development, progressing from Grace to Vera processors through extreme codesign optimized for LANL's simulation workloads. All three supercomputers were codesigned by hardware architects, system software developers, domain scientists, computer scientists and applied mathematicians to ensure the systems address real scientific workloads rather than abstract benchmarks alone.