130 parts · 1,202 companies
Prices as of2026-10-25
Search

Aurora: every supplier in the machine

Argonne National Laboratory, United States

Accelerator
63,744 Intel Data Center GPU Max 1550 (Ponte Vecchio), 6 per node, 128 GB HBM2e each
Host CPU
21,248 Intel Xeon CPU Max (Sapphire Rapids with HBM), 2 per node, 52 cores and 64 GB HBM2e each
Integrator
HPE (HPE Cray EX), built with Intel
Nodes
10,624 nodes (compute blades) in 166 cabinets, 64 blades per cabinet, eight rows
Memory
Per node: 1 TB DDR5 (2 x 512 GB), 128 GB HBM2e on the CPUs (2 x 64 GB), 768 GB HBM2e on the GPUs (6 x 128 GB). System: 10.62 PB DDR5, 9.52 PB HBM2e
Interconnect
HPE Slingshot-11: 8 Cassini 200 Gb/s NICs per node, 1-D dragonfly of 175 groups (166 compute, 8 storage, 1 service), 84,992 compute endpoints
Storage
DAOS on 1,024 server nodes, about 260 PB raw (about 220 PB with the suggested erasure coding), >=31 TB/s; plus the Flare Lustre file system, 91 PB
Cooling
100% direct liquid cooling inside the HPE Cray EX cabinets; Argonne facility loop with chillers, cooling towers, heat exchangers and filtration (44,000 gallons)
Power
38.7 (38,698.36 kW, TOP500 power measurement level 2; same value on the June 2024 through June 2026 lists)
Rmax (HPL)
1,012.00 PFlop/s (TOP500 June 2026 list)
Rpeak
1,980.01 PFlop/s (TOP500 June 2026 list)
Status
Final blade installed June 2023; first TOP500 entry November 2023; released for general research use in January 2025

Unmarked companies come from the operator's or a vendor's own pages (anl.gov, alcf.anl.gov, hpe.com, intel.com, the Argonne/Intel/HPE architecture paper, top500.org); 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: Argonne/Intel/HPE's own Aurora architecture paper (arXiv 2509.08207v2), ALCF user guides, HPE and Intel press materials and the TOP500 system page, plus one IEEE Spectrum article, researched 2026-10-02

This page is the parts-first view. For Aurora's full system record on the sister site Compute Atlas, an open dataset of supercomputer and AI-cluster supply chains, see Aurora on Compute Atlas.

  1. HPE Cray EX liquid-cooled cabinet (166 cabinets; 8 compute chassis and 8 switch chassis each, no fans)
    Rack-scale systems 4.01 · Chassis, rails, rack enclosures 4.14
    HPE (system integrator; Cray EX cabinet, chassis and Slingshot switch blades), Intel (named with HPE as builder of Aurora)
    Sources: anl.gov · arxiv.org · assets.ext.hpe.com
  2. Exascale Compute Blade (2x Xeon CPU Max + 6x GPU Max + 8 NICs; about 4 kW sustained, 145 TF/s FP64 in single-node HPL)
    GPU baseboards & compute trays 4.03
    HPE (Cray EX blade form factor and chassis), Intel (TOP500 names the node an "Intel Exascale Compute Blade") · which company builds the blade board and does the blade assembly (no source read names an ODM or EMS), PCIe switch vendor (the NICs hang off a PCIe switch on the blade; no vendor named)
    Sources: arxiv.org · top500.org
  3. Blade and cabinet power (rectifiers to 380 V DC, 8 on-blade IVOC converters from 380 V to 12 V)
    Power shelves & PSUs 4.12 · Rack PDUs 4.16
    HPE (cabinet PDUs and rectifiers are part of the Cray EX cabinet) · makers of the rectifiers and the IVOC converters (no source read names them)
    Sources: arxiv.org · assets.ext.hpe.com
  4. Intel Data Center GPU Max 1550 "Ponte Vecchio" (47 tiles, 2 stacks, 128 GB HBM2e; compute tiles on TSMC N5, base and cache tiles on Intel 7, Foveros + Co-EMIB packaging)
    Merchant GPUs / accelerators 1.01 · Leading-edge foundry 1.09 · 2.5D/3D advanced packaging 1.11 · Interposers, silicon bridges, RDL 1.16
    Intel (designer (Intel's own launch release); the Intel 7 base and cache tiles and Foveros / Co-EMIB packaging are per IEEE Spectrum), TSMC* (N5 compute tiles and N7 Xe Link tile) · HBM2e stack supplier (the paper, ALCF docs and Intel's launch release give the HBM2e capacity but name no memory maker)
    Sources: spectrum.ieee.org · download.intel.com · arxiv.org
  5. Xe Link GPU-to-GPU fabric (all-to-all among the 6 GPUs in a node, 28 GB/s per link)
    Scale-up fabric 3.01
    Intel (Xe Link is part of the GPU package), TSMC* (Xe Link tile fabricated on N7)
    Sources: arxiv.org
  6. Intel Xeon CPU Max "Sapphire Rapids" host CPU (52 cores, 4 tiles joined by EMIB, 64 GB on-package HBM2e, 350 W)
    Server CPUs 1.05 · 2.5D/3D advanced packaging 1.11
    Intel · DDR5 DIMM supplier (512 GB per socket; no source read names a memory maker)
    Sources: arxiv.org
  7. HPE Cassini NIC (200 Gb/s, 2 ports per NIC, 8 NICs per node, PCIe Gen4 host interface)
    NICs / SuperNICs 1.06
    HPE (designed the Cassini NIC ASIC) · foundry for the Cassini ASIC (not stated in any source read)
    Sources: arxiv.org
  8. Slingshot Rosetta switch ASIC (64 ports, 685 mm2, TSMC 16 nm FinFET, about 160 W typical)
    Scale-out switch ASICs 3.02 · Switch systems & white-box ODMs 3.03 · Leading-edge foundry 1.09
    HPE (Slingshot switch ASIC and the liquid-cooled switch blades in each cabinet), TSMC (fabricates Rosetta on 16 nm)
    Sources: arxiv.org · hpe.com
  9. Slingshot-11 dragonfly fabric and cabling (175 groups, one group per compute cabinet, electrical inside compute cabinets)
    Scale-out switch ASICs 3.02 · Copper: DAC / ACC / AEC cables 3.09
    HPE (Slingshot-11 fabric, topology and cabling) · cable and optical transceiver vendors (the paper says intra-group cables in compute cabinets are electrical and storage groups mix electrical and optical; no maker is named)
    Sources: arxiv.org
  10. DAOS storage servers (1,024 Intel Coyote Pass nodes, each 2x Ice Lake Xeon, 16x Optane PMem 200, 2x HPE Slingshot NICs)
    AI storage systems 2.10
    Intel (Coyote Pass server platform, Ice Lake CPUs and Optane persistent memory; DAOS software originated at Intel), HPE (Slingshot NICs in each DAOS server) · DDR4 DIMM supplier in the DAOS servers (32 GB DIMMs; no maker named)
    Sources: docs.alcf.anl.gov · arxiv.org
  11. DAOS flash tier (16x 15.3 TB Samsung PM1733 NVMe per server, 16,384 drives in total)
    Enterprise SSDs & controllers 2.08 · Drives 2.08a
    Samsung Electronics (PM1733 NVMe SSDs)
    Sources: docs.alcf.anl.gov
  12. Flare Lustre file system (HPE E1000-based, 650 GB/s peak, reached through gateway nodes)
    AI storage systems 2.10
    HPE (E1000 Lustre storage system)
    Sources: arxiv.org · docs.alcf.anl.gov
  13. In-cabinet liquid cooling (cold plates, manifolds, quick-disconnect blades, one CDU per up to four cabinets, facility water up to 32 C)
    Direct-to-chip cold plates 5.01 · Coolant distribution units 5.05
    HPE (Cray EX cabinet cooling loop and CDU) · makers of the cold plates, quick disconnects, CDU pumps and heat exchangers (HPE's brochure and Argonne's articles name none), Aurora-specific CDU count and the facility-water supply temperature actually used at Argonne
    Sources: assets.ext.hpe.com · alcf.anl.gov

No source read names the HBM2e or DDR5 memory suppliers, the PCIe switch, cable or optics vendors, or any maker of the chillers, cooling towers, heat exchangers and cold plates, so none are listed. The in-rack cooling and power lines rest on HPE's generic Cray EX cabinet brochure, which the Aurora paper cites for its cabinets; no Aurora-specific CDU or rectifier part list was found. The ALCF user guide and the architecture paper disagree on the DAOS and Lustre capacities, and both are noted in the file. 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:El Capitan · Frontier · 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.