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

Eos: every supplier in the machine

NVIDIA, United States

Accelerator
NVIDIA H100 Tensor Core GPU (SXM5, 80 GB HBM3); 8 per DGX H100 system, 4,608 in the 576-system build
Host CPU
2x Intel Xeon Platinum 8480C (56 cores each, 2.0 GHz base, 3.8 GHz max turbo) per DGX H100 system; 1,152 CPUs across 576 systems
Integrator
NVIDIA (DGX SuperPOD built from its own DGX H100 systems); TOP500 lists the manufacturer as Nvidia
Nodes
576 DGX H100 systems (4,608 GPUs). The TOP500 entry lists 485,888 cores; The Register reports the TOP500 run used 3,328 of the H100 GPUs
Memory
Per DGX H100 system: 640 GB HBM3 (8x 80 GB) plus 2 TB DDR5 system memory (32 DIMMs)
Interconnect
NVIDIA Quantum-2 InfiniBand NDR (up to 400 Gb/s); eight ConnectX-7 InfiniBand adapters per DGX H100 for the compute fabric; NVIDIA states 448 Quantum switches in a 3-tier, non-blocking, rail-optimized fabric; NVSwitch only inside each system
Storage
DDN EXAScaler AI storage (A3I): 48 appliances, 12 PB of flash, 4 TB/s, in under three racks at 100 kW
Rmax (HPL)
121.40 PFlop/s (TOP500 June 2026 list)
Rpeak
188.65 PFlop/s (TOP500 June 2026 list)
Status
Entered the TOP500 in November 2023 (TOP500 installation year 2023); NVIDIA's own internal AI development system, shown publicly in a video and blog post in February 2024; a separate 10,752-GPU Eos system used for MLPerf training is a different machine and is not part of this page

Unmarked companies come from NVIDIA's own pages (blogs.nvidia.com, nvidianews.nvidia.com, docs.nvidia.com, developer.nvidia.com), a vendor's own press release (ddn.com via PR Newswire), or top500.org; reported companies come from named trade press or analyst reporting 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: NVIDIA's Eos blog post (February 2024), its DGX H100 announcement, DGX H100 user guide, DGX SuperPOD H100 reference architecture and Hopper architecture write-up, DDN's press release on its Eos storage, and the TOP500 entry (June 2026 list), researched 2026-10-10

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

  1. System design and integration (NVIDIA DGX SuperPOD built from 576 DGX H100 systems, designed and assembled by NVIDIA for its own use)
    Rack-scale systems 4.01
    Nvidia (designs the DGX SuperPOD architecture and built Eos from its own DGX H100 systems; NVIDIA also states Eos serves as a blueprint for its OEM and cloud partners)
    Sources: blogs.nvidia.com · nvidianews.nvidia.com · top500.org · nvidia.com
  2. DGX H100 compute system (8U server with eight H100 GPUs on an NVSwitch baseboard, two Xeon hosts, 640 GB of HBM3 and 2 TB of DDR5, up to 10.2 kW)
    GPU baseboards & compute trays 4.03
    Nvidia (designs and sells the DGX H100 system; Eos is made of 576 of them) · maker of the six 3.3 kW power supplies, the DDR5 DIMMs, the NVMe drives (8x 3.84 TB data, 2x 1.92 TB boot) and the system boards (not named in the user guide or any Eos source)
    Sources: docs.nvidia.com · blogs.nvidia.com
  3. H100 GPU and in-system NVLink/NVSwitch (900 GB/s between any two GPUs in a DGX H100)
    Merchant GPUs / accelerators 1.01 · Scale-up fabric 3.01
    Nvidia (H100 Tensor Core GPU, fourth-generation NVLink and third-generation NVSwitch inside each DGX H100; NVIDIA states there is no NVSwitch fabric between nodes on Eos)
    Sources: docs.nvidia.com · nvidianews.nvidia.com · nextplatform.com
  4. H100 die fabrication (GH100, TSMC 4N process customised for NVIDIA)
    Leading-edge foundry 1.09
    TSMC (NVIDIA states the GH100 die that powers the H100 is fabricated on TSMC 4N; NVIDIA's Eos pages do not repeat the foundry) · advanced-packaging flow and substrate suppliers for the H100 SXM5 module (none named in any source read)
    Sources: developer.nvidia.com
  5. GPU memory (80 GB HBM3 per H100, five stacks; 640 GB per DGX H100)
    HBM 2.01
    SK Hynix* (reported as the primary HBM3 supplier for Nvidia's H100; no source read confirms it for the H100s in Eos) · which HBM3 vendors supplied the stacks on the H100s installed in Eos (NVIDIA does not say; TrendForce reports Samsung entered NVIDIA's supply only in late 2023 and in small volume)
    Sources: trendforce.com · developer.nvidia.com
  6. Host CPUs (two 56-core Xeon Platinum 8480C per DGX H100, 1,152 in total)
    Server CPUs 1.05
    Intel (4th-generation Xeon Scalable (Sapphire Rapids) host processors)
    Sources: docs.nvidia.com · top500.org · docs.nvidia.com (2)
  7. Network adapters (eight single-port ConnectX-7 InfiniBand cards for the compute fabric, two dual-port ConnectX-7 cards for storage and in-band management, per DGX H100)
    NICs / SuperNICs 1.06
    Nvidia (ConnectX-7 adapters, 400 Gb/s per port)
    Sources: docs.nvidia.com
  8. InfiniBand fabric (Quantum-2 NDR switches, 448 in a 3-tier, rail-optimized, non-blocking fat tree per NVIDIA's statement)
    Scale-out switch ASICs 3.02 · Switch systems & white-box ODMs 3.03
    Nvidia (Quantum-2 InfiniBand switches (the former Mellanox line); the switch count is NVIDIA's statement as quoted by The Next Platform) · optical transceiver, DAC and fibre cable vendors (NVIDIA publishes no cable bill of materials for Eos)
    Sources: blogs.nvidia.com · nextplatform.com · docs.nvidia.com
  9. Shared storage (DDN EXAScaler A3I, 48 appliances, 12 PB flash, 4 TB/s, under three racks at 100 kW)
    AI storage systems 2.10
    DDN (DDN states it implemented its EXAScaler AI storage in Eos; NVIDIA's DGX vice president is quoted in the same release saying the A3I delivers the storage for Eos) · NVMe drive vendor inside the A3I appliances (not named), whether DDN also supplies the separate user-storage tier the DGX SuperPOD reference architecture calls for (the release describes one 12 PB system)
    Sources: prnewswire.com
  10. Cluster software (NVIDIA Base Command and NVIDIA AI Enterprise)
    Cluster management & GPU orchestration 11.02
    Nvidia (Base Command for cluster management and AI Enterprise for the software stack)
    Sources: blogs.nvidia.com

NVIDIA does not publish a site address, a power draw or a cooling design for Eos. The TOP500 entry gives the site as NVIDIA Corporation, United States, and lists no power figure. No source read names the maker of the 2 TB of DDR5 DIMMs in each DGX H100, the NVMe drives (eight 3.84 TB data drives and two 1.92 TB boot drives per system), the six 3.3 kW power supplies, the system boards, the OSFP transceivers, DAC and fibre cables on the InfiniBand fabric, the racks, power distribution, cooling plant or data hall equipment. A February 2024 ServeTheHome article that read NVIDIA's video notes four DGX H100 systems per rack and liquid-cooled rack manifolds behind a cart in the footage, but it presents the rack power as an assumption and names no cooling supplier, so no cooling or rack line is made. NVIDIA's DGX SuperPOD reference architecture requires a second, Ethernet-attached user-storage tier; no source says what Eos uses for it, and the reference architecture's partner storage list is not tied to Eos. The TOP500 run used only part of the machine (The Register reports 3,328 of the 4,608 GPUs; NVIDIA has not published the figure), so the 121.40 PFlop/s figure is not the full machine's result. NVIDIA's March 2022 announcement described an Eos with 360 NVLink Switch System switches and about 275 PFlop/s of FP64 peak; that design was not built (see the interconnect spec). 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 · Aurora · JUPITER · Alps · LUMI · Leonardo · Isambard-AI · Venado · Perlmutter · Eagle (Microsoft Azure) · Fugaku · MareNostrum 5 · Horizon · Doudna (NERSC-10). See also the AI rack teardowns and thesupercomputer overview.