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

MareNostrum 5: every supplier in the machine

Barcelona Supercomputing Center, Spain

Accelerator
NVIDIA H100 (Hopper) with 64 GB HBM2e, 4 per ACC node, 1,120 ACC nodes (4,480 GPUs, derived); GPU pairs joined by NVLink at 300 GB/s bidirectional
Host CPU
ACC: 2x Intel Xeon Platinum 8460Y+ (Sapphire Rapids, 40 cores, 2.3 GHz) per node. GPP: 2x Intel Xeon Platinum 8480+ (56 cores, 2 GHz) per node; 72 GPP nodes use Intel Xeon CPU Max 9480 with HBM
Integrator
Bull SAS (Atos, now Eviden) as contractor, combining BullSequana XH3000 and Lenovo ThinkSystem SD650 V3 hardware; TOP500 lists the ACC entry as Bull and the GPP entry as Lenovo
Nodes
ACC: 1,120 nodes in 35 racks (32 per rack). GPP: 6,192 standard nodes (256 GB), 216 large-memory nodes (1 TB) and 72 HBM nodes, 6,480 in total (derived), in 89 racks plus 1 HBM rack, 72 nodes per rack
Memory
ACC node: 512 GB DDR5 (16 x 32 GB at 4.8 GHz) plus 4 x 64 GB HBM2e. GPP node: 256 GB DDR5 (216 nodes at 1 TB); the 72 HBM nodes carry 128 GB HBM2 plus 32 GB DDR5. System total not published
Interconnect
InfiniBand NDR200 in a three-layer fat tree with 324 QM9790 switches (3 GPP islands, 1 storage island, 7 ACC islands); GPP 1 x NDR200 shared by 2 nodes, ACC 4 x NDR200 per node (one per GPU)
Storage
IBM Storage Scale (GPFS) on IBM ESS-3500: 248 PB net / 376 PB gross HDD tier (18 TB NL-SAS) and 2.8 PB net / 4.8 PB gross flash tier (15.36 TB NVMe), 1.6 TB/s read and 1.2 TB/s write; IBM TS4500 tape archive of 400 PB
Cooling
Direct liquid cooling of compute nodes and power supplies on a 32-42 C circuit; a second 17-27 C circuit feeds rear-door 'cold doors' for air-cooled parts; ACC nodes are fully direct liquid cooled
Power
ACC: 4.16 MW (4,158.90 kW, level 2; 3.63 MW in the optimized run). GPP: 5.75 MW (5,752.90 kW, level 1). Measured during Linpack, TOP500 June 2026 list; no facility total published
Rmax (HPL)
ACC: 175.30 PFlop/s. GPP: 40.10 PFlop/s (TOP500 June 2026 list, two separate entries)
Rpeak
ACC: 249.44 PFlop/s. GPP: 46.37 PFlop/s (TOP500 June 2026 list). BSC quotes 260 PFlop/s (ACC), 46.4 PFlop/s (GPP) and 315.2 PFlop/s in total
Status
ACC and GPP partitions deployed and available to users in the second quarter of 2024 (TOP500 gives installation year 2023, first listed November 2023); co-funded by the EuroHPC JU with Spain, Portugal and Turkey, half the capacity going to EuroHPC users; next-generation partitions still in progress

Unmarked companies come from the operator's own documents (BSC-authored presentation and paper), the EuroHPC JU or the vendor's own documentation (nvidia.com); reported companies come from named trade press whose article body was read in full, not just the headline. No company on this page is marked reported. This is a map of who plays where, not a confirmed bill of materials: BSC's October 2025 MareNostrum 5 system-and-support presentation, the BSC-authored arXiv paper on the system (March 2025), the EuroHPC JU procurement announcement (May 2022), NVIDIA's Hopper architecture write-up and Quantum-2 switch manual, and the two TOP500 entries (June 2026 list), researched 2026-10-22

This page is the parts-first view. For MareNostrum 5's full system record on the sister site Compute Atlas, an open dataset of supercomputer and AI-cluster supply chains, see MareNostrum 5 on Compute Atlas.

  1. System design, contract and rack platforms (Bull SAS / Eviden BullSequana XH3000 for the accelerated partition, Lenovo ThinkSystem for the general-purpose partition)
    Rack-scale systems 4.01
    Eviden/Atos (contractor selected by the EuroHPC JU (as Bull SAS, Atos); BullSequana XH3000 platform for the ACC partition), Lenovo (ThinkSystem SD650 V3 platform for the GPP partition) · building, power feed and cooling-plant contractors for the new BSC hosting facility (EuroHPC describes it only as a new facility designed for the machine, run on green energy with heat reuse)
    Sources: eurohpc-ju.europa.eu · arxiv.org · indico.fccn.pt
  2. Accelerated-partition compute node (Eviden-designed: 2 Xeon 8460Y+, 4 H100 on NVLink, 512 GB DDR5, 480 GB NVMe; 1,120 nodes in 35 racks)
    GPU baseboards & compute trays 4.03
    Eviden/Atos (node designed by Atos R&D labs in France primarily for MareNostrum 5, per BSC's paper) · maker of the node's DDR5 DIMMs (16 x 32 GB) and 480 GB NVMe drive — not named in any source read
    Sources: arxiv.org · indico.fccn.pt
  3. General-purpose compute node (Lenovo ThinkSystem SD650 V3 dual-node trays, 12 nodes per chassis, 72 nodes per rack; 6,480 nodes in total)
    Motherboards, CPU trays, switch trays 4.04
    Lenovo (ThinkSystem SD650 V3 trays and chassis) · maker of the DDR5 DIMMs (16 GB and 64 GB) and of the 960 GB node-local NVMe drives — not named in any source read
    Sources: arxiv.org · top500.org
  4. H100 GPU (Hopper, 64 GB HBM2e, four per accelerated-partition node)
    Merchant GPUs / accelerators 1.01
    Nvidia (H100 GPUs; BSC's slides list Nvidia as a technology provider) · maker of the HBM2e stacks on the 64 GB H100 modules — BSC and Nvidia give the capacity but no memory supplier
    Sources: arxiv.org · indico.fccn.pt
  5. Hopper GPU die fabrication (TSMC 4N process)
    Leading-edge foundry 1.09
    TSMC (fabricates the GH100 die on its 4N process, per Nvidia's Hopper architecture write-up; BSC's documents do not name the foundry) · whether the 64 GB H100 module in MareNostrum 5 is packaged differently from the 80 GB parts Nvidia's write-up describes (BSC gives only the 64 GB HBM2e figure)
    Sources: developer.nvidia.com
  6. In-node GPU fabric (NVLink between the four H100 GPUs, 300 GB/s bidirectional between pairs; x16 PCIe Gen5 to the host CPUs)
    Scale-up fabric 3.01
    Nvidia (NVLink on the H100 modules)
    Sources: arxiv.org
  7. Host CPUs (Intel Sapphire Rapids: Xeon Platinum 8460Y+ in the accelerated partition, 8480+ in the general-purpose partition, Xeon Max 9480 in the 72 HBM nodes)
    Server CPUs 1.05
    Intel (all host and general-purpose processors) · maker of the 128 GB HBM2 stacks on the Xeon Max 9480 nodes — not named in any source read
    Sources: arxiv.org · indico.fccn.pt
  8. InfiniBand adapters (ConnectX-7 NDR200, one per GPU in the accelerated nodes; one per two nodes in the general-purpose partition)
    NICs / SuperNICs 1.06
    Nvidia (ConnectX-7 InfiniBand adapters, named in BSC's paper; BSC's slides list Nvidia as a technology provider)
    Sources: arxiv.org
  9. InfiniBand NDR200 fabric (324 QM9790 switches in a three-layer fat tree; 3 GPP islands, 7 ACC islands, 1 storage island)
    Scale-out switch ASICs 3.02 · Switch systems & white-box ODMs 3.03
    Nvidia (QM9790 is an NVIDIA Quantum-2 NDR InfiniBand switch model, per Nvidia's switch manual; BSC's paper names the model but not its maker) · optical transceiver, active optical cable and copper (DAC/ACC) cable vendors — BSC's network diagram distinguishes AOC/optic and DAC/ACC links but names no manufacturer, maker of the SN3700C top-of-rack switch (model named in BSC's paper without a maker) — not attributed
    Sources: arxiv.org · networking-docs.nvidia.com · indico.fccn.pt
  10. Parallel file system and flash/disk storage (IBM Storage Scale on ESS-3500 building blocks; 248 PB net HDD tier, 2.8 PB net flash tier)
    AI storage systems 2.10
    IBM (5th-generation Elastic Storage Server (ESS-3500) building blocks running Storage Scale (GPFS)) · maker of the 18 TB NL-SAS hard drives and of the 15.36 TB NVMe drives — not named in any source read, each ESS-3500 server contains a 48-core AMD EPYC Rome processor, per BSC's paper, but BSC lists no AMD role as a supplier, so AMD is not attributed on this line
    Sources: arxiv.org · indico.fccn.pt
  11. Tape archive (two IBM TS4500 libraries, 64 TS1160 drives, 400 PB on 20 TB JE cartridges)
    AI storage systems 2.10
    IBM (TS4500 libraries and TS1160 tape drives, driven by Spectrum Archive servers)
    Sources: arxiv.org · indico.fccn.pt
  12. In-rack liquid cooling and rear-door cooling (direct liquid cooling on a 32-42 C warm-water circuit, rear-door cold doors on a 17-27 C circuit)
    Direct-to-chip cold plates 5.01 · Rear-door heat exchangers 5.07
    Eviden/Atos (all ACC node components are direct liquid cooled), Lenovo (direct liquid cooling of the SD650 V3 compute nodes and power supplies) · makers of the cold plates, quick disconnects, coolant distribution units and rear-door heat exchangers — not named in any source read, the facility plant behind the two water circuits (dry coolers, chillers, heat-reuse interface) — not described beyond the supply temperatures
    Sources: arxiv.org · indico.fccn.pt

BSC names Eviden, Lenovo, IBM, Nvidia and Intel as the technology providers but does not name the makers of the DDR5 DIMMs, the HBM2e on the H100 modules, the HBM2 on the Xeon Max nodes, the NVMe drives in the nodes, the hard drives and flash drives behind the file system, the InfiniBand optics and copper cables, the cold plates, or the rear-door heat exchangers, so none of those get a line. The hosting building, its power feed and its cooling plant are described only in general terms (green energy, heat reuse, two water circuits), with no contractor or equipment maker named in any source read. The two TOP500 entries and BSC's own peak figures also differ (see the specs), and the next-generation partitions are still being built. 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) · Eos · Fugaku · Horizon · Doudna (NERSC-10). See also the AI rack teardowns and thesupercomputer overview.