LANL's New Vera-Powered Systems Deliver 7x Faster AI Agents Than Crossroads

Los Alamos National Laboratory's early benchmarks show its new Vera CPU beating the Crossroads x86 supercomputer by 7x on the URSA scientific AI agent framework.

That's not a synthetic peak-FLOPS number. On a concrete workload - an AI agent that forms hypotheses, chooses tools, launches simulations, analyzes outputs, and refines the next step - the Vera CPU crushes the older x86 chips. LANL built URSA (Universal Research and Scientific Agent) to do exactly that, running on Venado today and heading to Mission and Vision.

7x Faster for Scientific AI Agents with Vera

Agentic AI for science demands more than raw compute. Agents need fast single-threaded logic, memory bandwidth, and low-latency communication. Vera's custom Olympus core, LPDDR5 memory, and on-chip fabric deliver. A single Vera CPU outperforms a single socket x86 CPU by more than 3x across the board, but on agentic workflows like URSA the gap widens to 7x.

Memory is the limiter in these workloads. Vera gives you more than 4x the memory per core and 6x the memory per node compared to the Crossroads x86 system. That lets agents hold larger state, cache simulation outputs, and iterate without hitting swap.

Vera CPU Outruns x86 on Simulation Workloads

The advantage isn't just for AI agents. LANL tested Vera on Branson, an open-source Monte Carlo heat transfer simulation. Vera beats the Crossroads CPUs by over 3x. That's a direct apples-to-apples on a codesigned workload - LANL and NVIDIA have been co-designing CPUs for more than a decade, from Grace to Vera, using extreme codesign for real simulation workloads.

All of LANL's supercomputers are shaped by domain scientists, computer scientists, and applied mathematicians, not abstract benchmarks. The Vera CPU results on Branson prove the codesign philosophy works.

Three Supercomputers, One Architecture: Mission, Vision, Veritas

Mission, Vision, and Veritas will be built with HPE Cray Supercomputing GX5000 architecture and the NVIDIA Vera Rubin platform. Mission gets 2,300 standalone Vera CPUs alongside NVIDIA Vera Rubin GPU nodes. Veritas gets roughly 1,150 standalone Vera CPUs. Both use NVIDIA Quantum-X800 InfiniBand networking. Vision serves as a general science resource for materials, nuclear science, energy modeling, and biomedical research.

Mission is the fifth Advanced Technology System in the NNSA's Advanced Simulation and Computing program, replacing Crossroads for classified national security workloads. Veritas runs the Laboratory Directed Research and Development program, testing these technologies for even larger future systems.

All three are expected operational in 2027. When they go live, scientists will have a hardware platform purpose-built for agentic AI - a framework that can brainstorm hypotheses, plan experiments, run simulations, analyze results, and loop back autonomously. That's the real prize: turning the scientific method into an automated, high-throughput pipeline.

Source: NVIDIA Vera CPU Opens the Way for Agentic Scientific AI at Los Alamos National Laboratory
Domain: blogs.nvidia.com