El Capitan: every supplier in the machine
Lawrence Livermore National Laboratory, United States
- Accelerator
- AMD Instinct MI300A APU, 4 per node (6 GPU dies with 228 compute units, 3 Zen 4 CPU dies with 24 cores, 4 I/O dies, 8 HBM3 stacks per package)
- Host CPU
- None separate: each MI300A carries 24 Zen 4 x86 cores (3 CCDs of 8 cores) on the same package
- Integrator
- HPE (HPE Cray EX; TOP500 lists the system as HPE Cray EX255a)
- Nodes
- 11,104 compute nodes in 87 HPE Cray EX ("Olympus") cabinets with 29 CDUs (LLNL Getting Started slides, March 2026); 4 MI300A per node
- Memory
- 512 GB HBM3 per node (4 x 128 GB, one pool shared by CPU and GPU dies; no separate DRAM), 5.3 TB/s peak bandwidth per APU
- Interconnect
- HPE Slingshot-11, dragonfly topology, 64-port 25.6 Tb/s switches, 4 x 200 Gb/s interfaces per node (100 GB/s injection)
- Storage
- Global Lustre file system, 401 PB and 2.6+ TB/s, on 42 E1000 storage racks; plus 696 HPE "Rabbit" near-node modules (21 PB, 31 TB/s) tiered to Lustre
- Cooling
- 100% fanless direct liquid cooling in the cabinets (HPE) with 29 CDUs; facility warm-water plant of six 3,000-ton cooling towers, 28,000 tons total, using ASHRAE W3 water with chillers largely avoided
- Power
- 29.68 MW (29,684.62 kW, power measurement level 2, TOP500 June 2026 list entry; identical on the November 2025 list)
- Rmax (HPL)
- 1,809.00 PFlop/s (TOP500 June 2026 list)
- Rpeak
- 2,821.10 PFlop/s (TOP500 June 2026 list)
- Status
- Dedicated at LLNL on 2025-01-09; first benchmarked on the November 2024 TOP500 list; TOP500 gives installation year 2024; first exascale system built for NNSA
Unmarked companies come from the operator's or vendor's own pages (llnl.gov, hpc.llnl.gov, hpe.com); reported companies come from named trade press whose article body was read in full, not just the headline. This is a map of who plays where, not a confirmed bill of materials: LLNL documents (hpc.llnl.gov hardware and file-system pages, LLNL Getting Started slides, S&TR December 2024, llnl.gov news), HPE's November 2024 El Capitan press release, the TOP500 system entry, and named trade press, researched 2026-10-02
This page is the parts-first view. For El Capitan's full system record on the sister site Compute Atlas, an open dataset of supercomputer and AI-cluster supply chains, see El Capitan on Compute Atlas.
- System design and cabinets (HPE Cray EX, 87 liquid-cooled cabinets, built with AMD and Red Hat)Rack-scale systems 4.01HPE (integrator and Cray EX cabinet supplier)Sources: llnl.gov · hpc.llnl.gov
- Compute blade and node (HPE Cray EX255a; a blade holds two nodes, a node holds 4 MI300A APUs)
- AMD Instinct MI300A APU (CPU, GPU, cache and HBM3 in one package; 24 Zen 4 cores, 228 CDNA 3 compute units, 128 GB HBM3)AMD · HBM3 stack supplier (LLNL, HPE and AMD materials read do not name one; trade press on the wider MI300 family is not specific to the MI300A)Sources: hpc.llnl.gov · str.llnl.gov
- MI300A dies fabricated by TSMC (N5 compute dies, N6 I/O and cache dies) and joined with TSMC SoIC and CoWoS packagingSources: nextplatform.com · spectrum.ieee.org
- Slingshot-11 fabric (HPE Rosetta switches, dragonfly topology, 64 ports at 200 Gb/s)HPE · cable and optics suppliers (copper and optical links are not itemised in any source read)Sources: hpc.llnl.gov · hpe.com
- Slingshot-11 network interfaces (4 x 200 Gb/s per node, one per APU, on PCIe 5.0 x16)NICs / SuperNICs 1.06Sources: hpc.llnl.gov
- Rabbit near-node storage (4U module: 16 SSDs plus one AMD EPYC storage processor, two per cabinet, each wired to 16 compute nodes)HPE (co-developed the Rabbit with LLNL), AMD (EPYC storage processor in each Rabbit) · SSD maker (LLNL names the drive count per module but not the drive vendor)Sources: hpc.llnl.gov
- Global Lustre file system (401 PB, 2.6+ TB/s; 42 E1000 storage racks, HPE Cray ClusterStor E1000 per 2019 reporting)AI storage systems 2.10HPE* (HPE (Cray) ClusterStor E1000 named as El Capitan's external storage in 2019; LLNL's current slides list "42 E1000" racks) · drive makers (the E1000 uses NVMe SSDs and disk drives per the 2019 reporting; no drive vendor named for El Capitan), whether the delivered E1000 racks match the 2019 plan beyond the product name (LLNL's slides give "E1000" and capacity but not the model line)Sources: hpc.llnl.gov · blocksandfiles.com
- In-cabinet liquid cooling (100% fanless direct liquid cooling, 29 coolant distribution units)HPE (fanless direct liquid cooling architecture) · cold plate, manifold, quick-disconnect and CDU component suppliers (HPE does not itemise them)Sources: hpe.com
No source read names the maker of the HBM3 stacks, the Rabbit SSDs, the Lustre drives, the cabinet cold plates or coolant distribution units, or the Slingshot cables and optics, and LLNL does not name the manufacturers of the cooling towers, the 115 kV substation transformers or the switchgear. Cabinet, node, power and peak-performance figures also differ between LLNL documents and TOP500, as noted in the specs. The facility side, per LLNL, is a 115 kV feed into two 40 MW substation transformers (85 MW of building capacity) and a warm-water cooling plant of six 3,000-ton towers; no line is given to either here because LLNL names no equipment maker for them. Power and cooling outside the cabinet sit in Layer 6 and Layer 7.
Machine-readable: Markdown ·JSON ·every line of every machine (CSV)
Other machines:Frontier · Aurora · JUPITER · Alps · LUMI · Leonardo · Isambard-AI · Venado · Perlmutter · Eagle (Microsoft Azure) · Eos · Fugaku · MareNostrum 5 · Horizon · Doudna (NERSC-10). See also the AI rack teardowns and thesupercomputer overview.