Los Alamos Deploys NVIDIA Vera CPUs for Agentic AI Science
Los Alamos National Laboratory is deploying three new supercomputers, Mission, Vision, and Veritas, built with HPE and NVIDIA hardware including the NVIDIA Vera CPU to accelerate scientific discovery and agentic AI research. Early testing shows the Vera CPU delivers 7x higher performance on URSA (Universal Research and Scientific Agent) workloads and over 3x performance on Monte Carlo simulations compared to the previous Crossroads x86 supercomputer. The systems, expected operational in 2027, will support classified national security work, fundamental science research, and testing of AI agents that can autonomously form hypotheses, run simulations, and refine experiments.
TL;DR
- Los Alamos National Laboratory is building three new supercomputers using NVIDIA Vera CPUs, Rubin GPUs, and HPE Cray architecture
- Vera CPU showed 7x performance improvement on URSA agentic AI workloads and 3x improvement on Monte Carlo heat transfer simulations versus Crossroads x86
- Mission and Vision systems expected operational in 2027, with Mission replacing Crossroads for classified national security workloads
- LANL's URSA framework demonstrates agentic AI for science, enabling AI agents to form hypotheses, plan experiments, run simulations and analyze results
Why It Matters
Agentic AI for scientific research remains largely experimental, and LANL's deployment of purpose-built hardware and frameworks signals a shift toward production-scale autonomous scientific discovery. The 7x performance gain on URSA workloads demonstrates that specialized CPU architecture can meaningfully accelerate AI agent execution, not just traditional HPC simulations. This work establishes a reference architecture for institutions seeking to operationalize AI agents in high-consequence scientific domains.
Business Impact
The Vera CPU's performance gains on both agent and simulation workloads validate NVIDIA's strategy of custom silicon for specific computational patterns beyond general-purpose GPU acceleration. HPE's selection for the supercomputer architecture and LANL's decade-long collaboration with NVIDIA on CPU design creates a competitive moat in the national security and scientific computing markets. Organizations building agentic AI systems will likely look to LANL's results as a benchmark for infrastructure requirements and performance expectations.
Key Implications
- Agentic AI workloads may require different hardware optimization strategies than traditional LLM inference, favoring CPU-GPU-networking codesign over GPU-centric approaches
- LANL's public work on URSA and the Vera CPU performance data provide a reference implementation for other national labs and research institutions planning agentic AI infrastructure
- The three-year timeline to operational systems (2027) suggests agentic AI for science is moving from research prototype to production deployment at scale
What to Watch
Monitor whether Mission and Vision deliver the projected performance gains when operational in 2027 and how LANL's URSA framework evolves to handle more complex scientific workflows. Watch for adoption of Vera CPUs by other national labs or research institutions, which would signal broader industry confidence in the architecture. Track whether NVIDIA and HPE announce similar codesigned systems for other customers, indicating whether this model scales beyond LANL's specific requirements.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
