El Capitan: New supercomputer is the fastest
A new AMD-powered US supercomputer, El Capitan, has taken the top spot in the Top500 rankings with over 2,700 petaflops of peak performance, largely to support Lawrence Livermore’s nuclear weapons stockpile stewardship and fusion research. Commenters weigh the ethics of investing such capability in nuclear simulation versus disarmament, while noting that high‑precision physics codes demand far more compute than typical AI workloads and that industry AI clusters may now rival or exceed national lab systems in raw throughput. There is also debate over whether public rankings still reflect true global leadership, given unreported systems in China, large tech companies, and classified government installations.
Purpose and Role of El Capitan
- Built at Lawrence Livermore National Laboratory using AMD MI300A APU-based nodes interconnected by HPE’s Slingshot.
- Officially used to model nuclear weapon performance, aging, and safety, replacing live tests under test-ban regimes.
- Also expected to support other HPC workloads like fusion research, genomics, and fundamental simulations.
Nuclear Weapons, Deterrence, and Ethics
- Some commenters are disturbed that leading-edge compute is driven by nuclear weapons work, especially amid geopolitical tensions and stalled disarmament.
- Others argue supercomputer simulations are preferable to live nuclear testing and are essential for stockpile stewardship and credible deterrence.
- Clarified that modern work focuses more on reliability, safety, aging, and variable-yield designs than on ever-higher explosive yields.
- Concern raised that some states might neglect stewardship, risking discovering “use-by dates” on warheads only in crisis.
Why Nuclear Simulations Need Massive Compute
- Simulations couple many demanding domains: radiation and neutron transport, hydrodynamics, plasma physics, high-temperature chemistry, and aging effects.
- Extremely small time scales (nanoseconds), extreme conditions (pressures, temperatures, plasmas), and 3D modeling needs drive complexity.
- Codes often run large ensembles (uncertainty quantification, sensitivity analysis).
- There is debate over whether they simulate down to subatomic particles; consensus in thread is that full per-particle modeling is infeasible and heavy approximations are required.
Hardware, Performance, and Precision
- El Capitan is a significant win for AMD in exascale HPC, contrasting with less successful competing efforts.
- Discussion on FP64 vs lower-precision compute: nuclear/HPC workloads need high precision, unlike LLM training, which tolerates FP16/FP8.
- AI training clusters (e.g., tens of thousands of H100s) may now exceed national labs in raw (low-precision) FLOPs, but workloads and metrics are not directly comparable.
Topology, Secrecy, and Alternatives
- Key differentiator of supercomputers is low-latency, high-bandwidth interconnects and specialized topologies; many scientific codes are tightly coupled and not “embarrassingly parallel.”
- Distributed volunteer projects (Folding@home, SETI@home) work for loosely coupled problems, but not for many nuclear/HPC simulations.
- Top500 list is seen as incomplete: Chinese labs and major tech companies often withhold benchmark submissions due to sanctions, secrecy, or lack of incentive.
Historical and Miscellaneous Notes
- Fast Fourier Transform and global seismometer networks were partly driven by nuclear test detection, with significant spillover benefits to geophysics.
- Nostalgia for earlier supercomputers, front-panel lights, and comparison of historical FLOP records with modern consumer devices.