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

Leonardo: every supplier in the machine

CINECA (Bologna Technopole), Italy

Accelerator
NVIDIA A100 SXM4 custom 64 GB HBM2e (Ampere, NVLink 3.0); 4 per Booster node, 13,824 in total
Host CPU
Booster: 1x Intel Xeon Platinum 8358 (Ice Lake, 32 cores, 2.6 GHz) per node. Data Centric (DCGP): 2x Intel Xeon Platinum 8480+ (Sapphire Rapids, 56 cores, 2.0 GHz) per node
Integrator
Eviden (Atos), BullSequana XH2000 cabinets; TOP500 lists the manufacturer as Bull
Nodes
4,992 compute nodes: Booster 3,456 (single-node X2135 blades) plus Data Centric 1,536 (three-node X2140 blades); 155 racks in total, 138 of them compute racks
Memory
Booster node: 512 GB DDR4-3200 plus 4x 64 GB HBM2e. Data Centric node: 512 GB DDR5-4800. CINECA's website states 2,800 TB of RAM for the system
Interconnect
NVIDIA Mellanox HDR InfiniBand (QM8700 switches, 823 in total), Dragonfly+ topology with 23 cells; Booster nodes 2x dual-port ConnectX-6 HDR100 (400 Gb/s per node), Data Centric nodes 1x HDR100 (100 Gb/s)
Storage
Fast tier 5.7 PB raw all-flash NVMe (31x DDN ES400NVX2); capacity tier 137.6 PB raw (31x DDN ES7990X with SS9012 expansion enclosures, 246x 18 TB SAS HDD each); metadata on 4x DDN SFA400NVX; Lustre file system
Cooling
Direct liquid cooling with warm water: water enters the racks at 36 C and leaves at 42 C, then returns to 36 C through adiabatic drycoolers (no refrigeration cycle); CINECA states over 95% of the heat is removed by liquid
Power
7.49 MW measured during the Linpack run (TOP500 June 2026 list)
Rmax (HPL)
241.20 PFlop/s (TOP500 June 2026 list)
Rpeak
306.31 PFlop/s (TOP500 June 2026 list)
Status
Booster module in full production from July 2023 and Data Centric module from February 2024 (TOP500 installation year 2023); LISA AI partition inaugurated 11 June 2026 and not part of the TOP500 entry

Unmarked companies come from the operator's own pages (leonardo-supercomputer.cineca.eu, hpc.cineca.it, docs.hpc.cineca.it), CINECA-authored papers and presentations, or a vendor's own pages (atos.net, nvidia.com); reported companies come from named trade press or analyst reporting whose article body was read in full, not just the headline. No line on this page needed a reported tag. This is a map of who plays where, not a confirmed bill of materials: CINECA's Leonardo website and HPC user documentation, the CINECA-authored architecture paper (Turisini, Amati, Cestari, arXiv 2307.16885), a CINECA facility presentation, EuroHPC JU and Eviden/Atos press releases, and the TOP500 entry (June 2026 list), researched 2026-10-03

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

  1. System design, integration and rack platform (BullSequana XH2000 cabinets, 155 racks of 2 tons each; 138 compute racks)
    Rack-scale systems 4.01
    Eviden/Atos (system integrator and cabinet designer; Eviden is the Atos high-performance computing business, listed by TOP500 under the name Bull)
    Sources: eurohpc-ju.europa.eu · arxiv.org · atos.net
  2. Compute blades (Booster: single-node X2135 'Da Vinci' GPU blade, fan-less and liquid-cooled; Data Centric: three-node X2140 CPU blade)
    GPU baseboards & compute trays 4.03
    Eviden/Atos (designs both blade types) · maker of the 15 kW direct-liquid-cooled power shelves that feed the XH2000 cabinets (the brochure shows them but names no supplier)
    Sources: leonardo-supercomputer.cineca.eu · arxiv.org
  3. A100 GPU (custom 64 GB SXM4 part, four per Booster node; GA100 die fabricated on TSMC 7nm N7)
    Merchant GPUs / accelerators 1.01 · Leading-edge foundry 1.09
    Nvidia (A100 Tensor Core GPU, NVLink 3.0 inside the node), TSMC (Nvidia states the Ampere GA100 die is fabricated on TSMC N7; Nvidia's write-up describes the standard A100, and the 64 GB variant is a CINECA-specific configuration of the same GPU family) · which HBM2e vendors supplied the four 16 GB stacks on each 64 GB A100 (not named by CINECA, Eviden or Nvidia), advanced-packaging flow and substrate suppliers for the A100 module (none named in any source read)
    Sources: arxiv.org · developer.nvidia.com
  4. Host CPUs (Booster: one 32-core Xeon Platinum 8358 per node; Data Centric: two 56-core Xeon Platinum 8480+ per node)
    Server CPUs 1.05
    Intel (Ice Lake and Sapphire Rapids Xeon Scalable processors)
    Sources: hpc.cineca.it · arxiv.org
  5. System memory (Booster 512 GB DDR4-3200 per node; Data Centric 512 GB DDR5-4800 per node)
    DRAM 2.02
    Micron (Eviden's 2022 release says Leonardo relies on Micron DDR5 DRAM; DDR5 is fitted only in the Data Centric nodes) · DRAM vendor for the Booster DDR4-3200 DIMMs (not named), which module maker or makers supplied the DDR5 DIMMs built from Micron DRAM (not named)
    Sources: atos.net · hpc.cineca.it
  6. Network adapters (ConnectX-6 HDR100: two dual-port cards per Booster node, one single-port card per Data Centric node)
    NICs / SuperNICs 1.06
    Nvidia (Mellanox ConnectX-6 InfiniBand adapters)
    Sources: arxiv.org
  7. InfiniBand fabric (QM8700 HDR switches, 823 in total, Dragonfly+ across 23 cells with an all-to-all link between every cell pair)
    Scale-out switch ASICs 3.02 · Switch systems & white-box ODMs 3.03
    Nvidia (Quantum QM8700 HDR switches (the former Mellanox line); the paper's description of the four Ethernet-to-InfiniBand gateways cites Nvidia's Skyway datasheet) · optical transceiver, DAC and fibre cable vendors (the paper gives link lengths of 1 m, 5 m and 20 m but names no cable maker)
    Sources: arxiv.org · docs.hpc.cineca.it
  8. Fast storage tier (5.7 PB raw all-flash, 31x DDN ES400NVX2 appliances with 24x 7.68 TB NVMe SSDs each)
    AI storage systems 2.10
    DDN (EXAScaler appliances and Lustre software) · NVMe SSD vendor inside the ES400NVX2 appliances (not named)
    Sources: leonardo-supercomputer.cineca.eu
  9. Capacity tier and metadata (137.6 PB raw: 31x DDN ES7990X controllers with SS9012 expansion enclosures, 246x 18 TB SAS HDD per module; 4x DDN SFA400NVX flash metadata units)
    AI storage systems 2.10
    DDN (ES7990X, SS9012 and SFA400NVX appliances; 66 DDN appliances in total across both tiers, in 12 racks) · maker of the 18 TB SAS hard drives and of the metadata NVMe SSDs (not named)
    Sources: leonardo-supercomputer.cineca.eu · arxiv.org
  10. In-cabinet liquid cooling (XH2000 direct liquid cooling on warm water: cold plate on every 1U blade, hydraulic chassis with heat exchanger, secondary manifolds, non-spill quick disconnects)
    Direct-to-chip cold plates 5.01 · Coolant distribution units 5.05
    Eviden/Atos (the direct-liquid-cooling system is part of the XH2000 cabinet design; the brochure describes the hydraulic chassis as transferring 95% of the heat between the primary and secondary loops) · makers of the cold plates, manifolds, quick-disconnect couplings and pump and heat-exchanger components inside the hydraulic chassis (not named by Eviden or CINECA)
    Sources: atos.net · arxiv.org
  11. Data hall UPS (four branches of 2 x 1 MW Vertiv Trinergy units with lithium batteries)
    UPS (static, modular, rotary/DRUPS) & UPS power semis 7.10
    Vertiv (Trinergy UPS units; CINECA's slide labels the installation "Vertiv UPS System") · maker of the lithium battery cells and cabinets (the slide says "Lithium Battery" and names no supplier)
    Sources: vertiv.com · it-online.co.za · leonardo-supercomputer.cineca.eu
  12. Frontend and service nodes (32 BullSequana X430-E6 and X450-E6 frontend servers on Xeon Platinum 8358 with 16 Nvidia Quadro RTX 8000 visualisation GPU pairs; 11 management nodes on AMD EPYC 7H12)
    Merchant GPUs / accelerators 1.01 · Server CPUs 1.05
    Eviden/Atos (frontend server models are BullSequana X430-E6 (login) and X450-E6 (graphical)), Intel (Xeon Platinum 8358 in the frontend nodes), AMD (EPYC Rome 7H12 (64 cores) in the 11 operational management nodes), Nvidia (two PCIe Quadro RTX 8000 48 GB per graphical node) · system maker of the management nodes (the paper lists the CPU and NICs but not the server model)
    Sources: arxiv.org
  13. System management and monitoring software (Bull Energy Optimizer, Bull Dynamic Power Optimizer, SMC xScale; ParaStation HealthChecker)
    Cluster management & GPU orchestration 11.02
    Eviden/Atos (Bull Energy Optimizer, Bull Dynamic Power Optimizer and the SMC xScale monitoring suite), ParTec (ParaStation HealthChecker for issue detection and tracking, and ParaStation container support)
    Sources: arxiv.org · leonardo-supercomputer.cineca.eu

CINECA, Eviden and EuroHPC name Eviden, Nvidia, Intel, DDN, Micron (for DDR5) and Vertiv, but no source read names the maker of the HBM2e stacks on the 64 GB custom A100, the DDR4 DIMMs on the Booster nodes, the 18 TB SAS hard drives or the NVMe drives inside the DDN appliances, the cold plates, manifolds and quick disconnects inside the XH2000 cabinets, the 15 kW cabinet power shelves, or the InfiniBand optics and cables. The data hall facts are published without supplier names, so they sit here and not in a line. The hall is a 32 x 23 x 3.8 m reinforced-concrete room without internal pillars, rated seismic class 4 and REI 240 for fire, with a raised floor, four cooling tunnels in the basement, four MV/LV power stations on the first floor, aluminium busbars (3,200 A primary, 1,000 A secondary) above the racks, and adiabatic drycoolers returning the warm-water loop to its supply temperature; the room air is conditioned with chilled water from refrigeration units. None of the transformers, busbars, drycoolers or chillers is attributed to a manufacturer in the pages read. LISA, the AI partition added in 2026 (166 eight-way GPU servers from Bull), is not part of the TOP500 entry or of the figures on this page, and its component suppliers were not researched. 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 · Isambard-AI · Venado · Perlmutter · Eagle (Microsoft Azure) · Eos · Fugaku · MareNostrum 5 · Horizon · Doudna (NERSC-10). See also the AI rack teardowns and thesupercomputer overview.