JUPITER: every supplier in the machine
Forschungszentrum Jülich, Germany
- Accelerator
- NVIDIA GH200 Grace Hopper Superchip (Hopper GPU, 96 GB HBM3 at 4 TB/s per GPU); 4 per Booster node, about 24,000 in total
- Host CPU
- Booster: NVIDIA Grace (72 Arm Neoverse V2 cores) in each superchip. Cluster: 582 nodes of 2x AMD EPYC Turin 9655 (96 cores) plus 162 nodes of 2x SiPearl Rhea1 (80 Arm Neoverse V1 cores)
- Integrator
- ParTec-Eviden consortium led by Eviden (now Bull), BullSequana XH3000 racks
- Nodes
- Booster: 5,884 nodes in about 125 racks. Cluster: 582 AMD Turin nodes plus 162 SiPearl Rhea1 nodes, still being installed
- Memory
- Booster node: 4x 120 GB LPDDR5X plus 4x 96 GB HBM3. Cluster: Turin nodes 768 GB DDR5 (a minority with about 1.5 TB); Rhea1 nodes 2x 64 GB HBM2e plus 512 GB DDR5. System total not published
- Interconnect
- NVIDIA Mellanox Quantum InfiniBand NDR, Dragonfly+ topology, 27 groups (25 Booster, 1 Cluster, 1 administrative); 4x NDR200 ConnectX-7 per Booster node, 1 per Cluster node; 867 switches
- Storage
- ExaFLASH 29 PB flash (about 20 PB usable, 20x IBM Storage Scale System 6000); ExaSTORE 308 PB raw disk (22x IBM Storage Scale System 6000); ExaTAPE 379 PB (2x IBM TS4500 libraries); IBM Storage Scale (GPFS) file system
- Cooling
- Warm-water direct liquid cooling in the BullSequana XH3000 racks; waste heat is to feed the Jülich campus heating network
- Power
- 15.79 MW measured during the Linpack run (TOP500 11/2025 list; the 06/2026 list repeats the same entry)
- Rmax (HPL)
- 1,000.00 PFlop/s (TOP500 June 2026 list)
- Rpeak
- 1,226.28 PFlop/s (TOP500 June 2026 list)
- Status
- Booster module inaugurated 5 September 2025 (TOP500 installation year 2025); Cluster module still being installed, first SiPearl Rhea1 samples delivered to Bull on 22 September 2026
Unmarked companies come from the operator's or a vendor's own pages (fz-juelich.de, nvidia.com, sipearl.com, par-tec.com, eiffage.com); 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: Jülich Supercomputing Centre's JUPITER technical overview and user documentation, vendor press releases (Nvidia, SiPearl, Eiffage) and top500.org, researched 2026-10-02
This page is the parts-first view. For JUPITER's full system record on the sister site Compute Atlas, an open dataset of supercomputer and AI-cluster supply chains, see JUPITER on Compute Atlas.
- System design, integration and rack platform (BullSequana XH3000, delivered by the ParTec-Eviden consortium)Eviden/Atos (consortium lead; BullSequana XH3000 racks and blades, integrated at its Angers factory; Eviden now trades as Bull), ParTec (consortium partner; dynamic Modular System Architecture and system software)
- Booster compute node (4x GH200 superchips, 4x NDR200 InfiniBand adapters, 960 GB NVMe boot disk; 5,884 nodes in about 125 racks)Sources: apps.fz-juelich.de · fz-juelich.de
- GH200 Grace Hopper Superchip (Hopper GPU plus 72-core Grace CPU joined by NVLink-C2C)Sources: fz-juelich.de · nvidianews.nvidia.com
- Hopper GPU die fabrication (TSMC 4N process)Leading-edge foundry 1.09TSMC (fabricates the GH100 GPU die on its 4N process, per Nvidia's Hopper architecture write-up; JUPITER's own pages do not name the foundry) · advanced-packaging flow and substrate suppliers for the GH200 module (none named in any source read)Sources: developer.nvidia.com
- GPU memory (96 GB HBM3 per Hopper GPU, 4 TB/s)HBM 2.01SK Hynix* (reported as the primary HBM3 supplier for Nvidia's H100-generation GPUs; no source read confirms it for the GH200 units installed at JUPITER) · which HBM3 vendors supplied the stacks on JUPITER's GH200 modules (the Jülich pages and Nvidia's press release do not say)Sources: trendforce.com · news.skhynix.com
- System memory (Booster 4x 120 GB LPDDR5X per node; Cluster 768 GB DDR5 per Turin node, 512 GB DDR5 per Rhea1 node)DRAM 2.02Nvidia (the Booster LPDDR5X sits on the Grace Hopper module), AMD (Turin nodes use DDR5 through the EPYC memory controllers) · DRAM vendor for the Booster LPDDR5X and the Cluster DDR5 DIMMs (not named by the operator or either vendor)Sources: apps.fz-juelich.de
- Cluster module CPUs, AMD EPYC Turin 9655 (582 dual-socket nodes, 96 cores per CPU)Server CPUs 1.05
- Cluster module CPUs, SiPearl Rhea1 (162 dual-socket nodes, 80 Arm Neoverse V1 cores and 64 GB HBM2e per CPU; first samples delivered to Bull 2026-09-22)SiPearl (designer of the Rhea1 CPU), Arm (Neoverse V1 core design), TSMC (SiPearl states Rhea1 is manufactured at TSMC) · HBM2e vendor for the four in-package stacks per Rhea1 (not named), final delivery and acceptance date for the Rhea1 partition (SiPearl says first samples only, with release at the end of 2026)
- InfiniBand adapters (ConnectX-7, NDR200: 4 per Booster node, 1 per Cluster node)NICs / SuperNICs 1.06Sources: apps.fz-juelich.de
- InfiniBand fabric (Quantum NDR switches, Dragonfly+ with 27 groups, about 50,000 links and 867 switches)Nvidia (Quantum-2 InfiniBand switches (the former Mellanox line)) · optical transceiver, DAC and fibre cable vendors for the roughly 50,000 links (not named in any source read)Sources: fz-juelich.de · nvidianews.nvidia.com
- Flash and disk storage tiers (ExaFLASH and ExaSTORE on IBM Storage Scale System 6000 building blocks) and tape archive (ExaTAPE, IBM TS4500 libraries)AI storage systems 2.10IBM (Storage Scale System 6000 appliances and software, TS4500 tape libraries, Storage Protect) · NVMe flash and hard-drive vendors inside the storage appliances (not named)Sources: fz-juelich.de
- Rack-level cooling (BullSequana XH3000 direct liquid cooling on warm water, 37-47 C circulating)Eviden/Atos (patented direct liquid cooling built into the XH3000 rack and blade design) · cold-plate, manifold, quick-disconnect and in-rack CDU component suppliers (not named by Eviden/Bull or Jülich)
- Modular data centre (about 50 factory-built containers on 2,300 m2: 20 IT, 15 energy supply, about 10 logistics; racks installed and tested before delivery)Eviden/Atos (delivered the container data centre under a separate contract signed January 2024), Eiffage (Eiffage Énergie Systèmes fitted the modules with electrical equipment and the hydraulic cooling networks)Sources: eiffage.com · fz-juelich.de
- Electrical distribution inside the modules (35 kV feed, fifteen 2,500 A electrical cabinets, UPS room)Eiffage (installed the fifteen 2,500 A electrical cabinets, distribution, cable trays and cabling) · switchgear, transformer and UPS manufacturers (Eiffage and Jülich describe the installation but do not name the equipment makers)Sources: eiffage.com · heise.de
- Heat reuse and heat rejection (waste water above 40 C piped to Jülich campus buildings; rooftop heat exchangers as backup)Eviden/Atos (free-cooling and direct-liquid-cooling design), Eiffage (hydraulic networks under the technical floor of each module) · manufacturer of the rooftop heat exchangers and the campus heat-network interface (not named)Sources: fz-juelich.de · globenewswire.com
- System management and modular system software (JUPITER Management Stack, ParaStation Modulo, xScale)Sources: fz-juelich.de
The Jülich pages name Nvidia, AMD, SiPearl, IBM and Eviden/Bull as suppliers but do not name the DRAM, HBM or flash vendors behind any of the memory or storage tiers, the optics and cable vendors for the InfiniBand fabric, or the UPS and cooling-plant manufacturers inside the container data centre. The Cluster module is still being installed, so its final node counts could change. 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 · 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.