Supercomputers
The AI data centers in the rest of this catalog and the national-laboratory supercomputers on the TOP500 list are built from the same part categories: accelerators, memory, a high-speed network, storage, liquid cooling, and the power behind them. This section takes 16 flagship machines and walks each one from the cabinet inward, the way the rack teardowns do for AI racks, with every line mapped to catalog parts and the companies named on them.
Placements marked with an asterisk come from named trade or analyst reporting rather than the operator's or vendor's own pages. A rank would go stale at the next TOP500 list, so none is stated here; the Rmax and power figures are as listed on the edition each page cites. Where a TOP500 entry lists no power, the table says so (Eagle, Eos) or gives the operator's design limit (Isambard-AI).
Compute Atlas is a sister site and an open dataset of the supply chain of supercomputers and AI clusters: which companies' accelerators, CPUs, interconnect, storage and cooling appear in which systems, each fact carrying a source link and a confidence tier. It tracks hundreds of systems. This section goes the other way, deep on a few machines and wired into the parts and companies of this catalog, so a supplier named here links straight to everything else it makes.
Side by side
| System | Site | Accelerator | Host CPU | Interconnect | Integrator | Power | Rmax (HPL) |
|---|---|---|---|---|---|---|---|
| El Capitan | Lawrence Livermore National Laboratory, United States | AMD Instinct MI300A APU | None separate (24 Zen 4 cores per APU) | HPE Slingshot-11 | HPE | 29.7 MW | 1,809 PFlop/s |
| Frontier | Oak Ridge National Laboratory, United States | AMD Instinct MI250X | AMD EPYC (64 cores) | HPE Slingshot-11 | HPE | 24.6 MW | 1,353 PFlop/s |
| Aurora | Argonne National Laboratory, United States | Intel Data Center GPU Max 1550 | Intel Xeon CPU Max (52 cores) | HPE Slingshot-11 | HPE, with Intel | 38.7 MW | 1,012 PFlop/s |
| JUPITER | Forschungszentrum Jülich, Germany | NVIDIA GH200 Grace Hopper | NVIDIA Grace; Cluster: AMD EPYC Turin, SiPearl Rhea1 | NVIDIA InfiniBand NDR | Eviden, with ParTec | 15.8 MW | 1,000 PFlop/s |
| Alps | CSCS (Swiss National Supercomputing Centre), Switzerland | NVIDIA GH200 Grace Hopper | NVIDIA Grace; AMD EPYC in other partitions | HPE Slingshot-11 | HPE | 7.1 MW | 435 PFlop/s |
| LUMI | CSC IT Center for Science (Kajaani), Finland | AMD Instinct MI250X | AMD EPYC Trento (64 cores) | HPE Slingshot-11 | HPE | 7.1 MW | 380 PFlop/s |
| Leonardo | CINECA (Bologna Technopole), Italy | NVIDIA A100 (64 GB) | Intel Xeon (Ice Lake, Sapphire Rapids) | NVIDIA HDR InfiniBand | Eviden (Atos) | 7.49 MW | 241 PFlop/s |
| Isambard-AI | University of Bristol (Bristol Centre for Supercomputing), United Kingdom | NVIDIA GH200 Grace Hopper | NVIDIA Grace (Arm Neoverse V2) | HPE Slingshot 11 | HPE | Under 5 MW | 217 PFlop/s |
| Venado | Los Alamos National Laboratory, United States | NVIDIA GH200 Grace Hopper | NVIDIA Grace (Arm Neoverse V2) | HPE Slingshot 11 | HPE | 1.66 MW | 99 PFlop/s |
| Perlmutter | NERSC (Lawrence Berkeley National Laboratory), United States | NVIDIA A100 | AMD EPYC 7763 (64 cores) | HPE Slingshot 11 | HPE | 2.9 MW | 79 PFlop/s |
| Eagle (Microsoft Azure) | Microsoft Azure, United States | NVIDIA H100 (SXM, 80 GB HBM3) | Intel Xeon Platinum 8480C | NVIDIA Quantum-2 InfiniBand NDR | Microsoft | Not published | 561 PFlop/s |
| Eos | NVIDIA, United States | NVIDIA H100 | Intel Xeon Platinum 8480C | NVIDIA Quantum-2 InfiniBand NDR | NVIDIA | Not published | 121 PFlop/s |
| Fugaku | RIKEN Center for Computational Science, Japan | None (CPU-only) | Fujitsu A64FX (Arm) | Tofu Interconnect D | Fujitsu | 29.9 MW | 442 PFlop/s |
| MareNostrum 5 | Barcelona Supercomputing Center, Spain | NVIDIA H100 (64 GB HBM2e) | Intel Xeon (Sapphire Rapids) | InfiniBand NDR200 | Bull SAS (Eviden), with Lenovo | 4.16 MW (ACC, Linpack) | 175 PFlop/s (ACC) |
| Horizon | Texas Advanced Computing Center (University of Texas at Austin), United States | NVIDIA Blackwell (GB200 NVL4) | NVIDIA Grace; Vera planned | NVIDIA XDR InfiniBand | Dell Technologies | About 13 MW (TACC estimate) | Not listed on TOP500 |
| Doudna (NERSC-10) | National Energy Research Scientific Computing Center, Lawrence Berkeley National Laboratory, United States | NVIDIA Rubin | NVIDIA Vera | NVIDIA Quantum-X800 InfiniBand | Dell Technologies | Not published | Not published |
The machines
Machine-readable: JSON index ·every line of every machine (CSV) · each machine also has a Markdown and a JSON version, linked at the foot of its page.
For AI-specific racks and pods see the rack teardowns; for named campuses see real facilities; to see how the parts connect seehow it fits together.