Supercomputers/Systems/El Capitan
USA · operational · classified
El Capitan
Deep dive: El Capitan: building the NNSA exascale system
Also known as ATS-4
Operated by Lawrence Livermore National Laboratory at Livermore Computing Center .
Researched with AI assistance, checked by scripts. These rows were researched by AI agents reading public operator, laboratory and vendor sources, appended by scripts and checked by automated validation; no subject-matter expert has reviewed them (how this site is made). A mechanical second-reader pass has since fetched every cited source: 1350 of 1540 systems have a readable citation that names them, and 64 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
Measured performance
- Rmax
- 1.81 EFlop/s
- Rpeak
- 2.82 EFlop/s
- Rmax ÷ Rpeak
- 64.1%
- Cores
- 11,340,000
- Power
- 29.68 MW
- Per watt
- 60.9 GF/W
Figures are as last publicly reported for the configuration described below, not a live measurement. Where a system was upgraded in place, the post-upgrade configuration is shown and the earlier one appears in the timeline.
What this machine is made of
One row per supplier relationship. “Supplier at build” is the company that shipped the part at the time; where that company has since been acquired, the parent it rolls up to today is shown beside it. That distinction is what makes ticker-level aggregation possible across a thirty-year dataset.
| Role | Part | Supplier at build | Quantity | Confidence | Source |
|---|---|---|---|---|---|
| Integrator | - HPE Cray EX255a | Hewlett Packard Enterprise | - | Verified | llnl.govhpe.com |
| Accelerator | AMD Instinct MI300A Accelerators ROCm (CDNA 3) · TSMC N5/N6 · 550 W AMD Instinct MI300A | AMD | - | Verified | llnl.govhpe.com |
| CPU | AMD 4th Gen EPYC 24C (in MI300A) CPUs x86-64 · 24C · 1.8 GHz · TSMC N5 AMD 4th Gen EPYC 24C 1.8GHz (integrated in MI300A) | AMD | - | Reported | llnl.govhpe.com |
| Interconnect | HPE Slingshot-11 Interconnect Dragonfly · 200 Gb/s per port · Around 2 microseconds Slingshot-11 | Hewlett Packard Enterprise | - | Verified | llnl.govhpe.com |
| Storage | HPE Cray ClusterStor E1000 Storage Lustre appliance, flash and disk tiers · 700 PB on Frontier (Orion) · About 10 TB/s peak Lustre on ClusterStor E1000 | Hewlett Packard Enterprise | - | Reported | llnl.govhpe.com |
| Cooling | Direct liquid cooling, warm water Cooling Direct-to-chip cold plates, facility water loop · Removes 90+ percent of node heat to liquid 100% direct liquid cooled cabinet | Hewlett Packard Enterprise | - | Reported | llnl.govhpe.com |
| Operating system | Tri-Lab Operating System Stack Operating systems Linux distribution, RHEL-derived, HPC-tuned · Open source, LLNL-maintained Tri-Lab Operating System Stack (TOSS 4), a RHEL-based distribution | Lawrence Livermore National Laboratory | - | Reported | hpc.llnl.gov |
AI datacenter metrics
Figures beyond an accelerator count, each with who stated it. What can and cannot be compared across AI datacenters.
| Metric | Value | As of | Basis | Note | Source |
|---|---|---|---|---|---|
| Accelerators | 45,696 accelerators | 2026-10 | Operator stated | LLNL job-limits page: 11,424 batch nodes x 4 MI300A APUs. Excludes 16 login and 56 debug nodes; other published counts use a different node total. | hpc.llnl.gov |
| IT power | 35 MW | 2026 | Operator stated | LLNL ASC page: peak power about 35 MW (same page also says about 30 MW at peak; 2022 LLNL news says 30-35 MW). Upper bound of the stated range. | asc.llnl.govllnl.gov |
Published benchmark results
- HPL-MxP 2025-11: HPL-MxP mixed-precision (rank 1), 16.68 EFlop/s (independent, 2025-11) hpl-mxp.orgtop500.org
- HPL-MxP 2026-06: HPL-MxP mixed-precision (rank 1), 16.7 EFlop/s (independent, 2026-06) top500.orgtop500.org
The read
Built on the MI300A APU, which collapses the CPU and accelerator roles onto one package with unified HBM3: there is no host-to-device copy. That has a direct consequence for this dataset: the CPU and accelerator edges point at parts made of the same silicon by the same supplier, and any accelerator-share analysis that counts CPU sockets separately will double-count. Our methodology page states how we handle it. Stockpile stewardship workload.
Timeline
- 2019 Announced source
- 2024-11 List debut at #1 (November 2024) source
- 2024-11 First operational source
Public list appearances
Pointers only: rank and edition, linking to the canonical entry. We do not reproduce list tables. See sourcing policy.
Site & facility
Livermore Computing Center
- Location
- Livermore, CA, USA
- Cooling
- Direct liquid cooling (warm water)
- Commissioned
- 1996
Change history
- 2026-10-04 AI benchmark record: hpl-mxp 2025-11
- 2026-10-04 AI benchmark record: hpl-mxp 2026-06
- 2026-09-22 Gained a Green500 rank pointer (#25 on the 2025-06 list)
Source check
We fetched this system's own citations and recorded whether each page actually mentions it. This is a corroboration signal, not a fact check, and it is published so you can see how well the sourcing holds up rather than take it on trust.
| Cited source | Result | Found on the page |
|---|---|---|
| hpe.com | does not name this system | Direct liquid cooling, warm water, HPE Cray ClusterStor E1000 |
| llnl.gov | names system and part | AMD Instinct MI300A, Direct liquid cooling, warm water, Tri-Lab Operating System Stack, HPE Cray ClusterStor E1000 |
Checked 2026-10-18 by pnpm verify. Re-run it and the table changes with the web.
Related news
- TOP500's 66th list confirms JUPITER as the first exascale system outside the US · TOP500El Capitan holds number one at a remeasured 1.809 exaflop/s and Frontier and Aurora hold two and three, so all three US exascale systems in this dataset keep their relative order. JUPITER's 1.000 exaflop/s Booster partition becomes the fourth exascale system worldwide and the first in Europe, confirming the milestone this dataset's earlier rows on Bull's JUPITER delivery had been building toward.
- Google's Ironwood (TPU v7) becomes generally available, compared directly to El Capitan · GoogleGoogle's own launch materials put a full Ironwood pod at 42.5 exaflops of FP8 compute, described as more than 24x El Capitan's 1.7 exaflops figure, though the two numbers measure different things, dense FP8 inference throughput versus double precision HPL, and are not a like for like comparison. It is nonetheless the most direct public benchmark claim yet tying a hyperscaler's own accelerator generation to a specific system already in this dataset.
- El Capitan verified as the fastest system in the world · Lawrence Livermore National LaboratoryLLNL confirms the MI300A-based system past 1.7 EFlop/s, making it the third US exascale machine and the second built entirely on AMD silicon.
Further reading
Inside a node
Why accelerators exist, what the memory hierarchy costs, and how the CPU and accelerator merged onto one package.
Power and cooling
Why air ran out at 20 kW a rack, how warm-water cooling removes the chillers, and what PUE conceals.
Operating a supercomputer
Acceptance testing, why "installed" and "in production" are six months apart, and the machine lifecycle.