COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-02

· Compute Atlas

JUPITER: Europe's exascale, American silicon

JUPITER hit 1.000 exaflop/s in November 2025 on a 500 million euro budget. The Booster's CPUs, GPUs, fabric and flash are American; the European CPU reached sample stage only in September 2026.

JUPITER is the first exascale system in Europe and the first outside the United States. It sits at Forschungszentrum Julich, it is owned by the EuroHPC Joint Undertaking, and on the TOP500 list of 17 November 2025 its Booster module measured 1.000 exaflop/s on HPL, fourth in the world. On the current top500.org system page it is fifth, at 1,000.00 PFlop/s Rmax, 1,226.28 PFlop/s Rpeak and 15,793.81 kW.

The number most worth knowing is not the exaflop. It is the split between what Europe paid for and what Europe made. The money is European: a 500 million euro budget shared by the EU and two German governments. The integration is European: Eviden’s BullSequana XH3000 platform, with ParTec’s software and system design. The processors, accelerators, network and flash are not. The one European processor in the design, SiPearl’s Rhea1, only began shipping samples to the integrator on 22 September 2026, more than a year after the machine was inaugurated.

How it came to exist

The sequence, from the sources we could fetch:

  • 14 December 2022: EuroHPC and Forschungszentrum Julich sign the hosting agreement. Julich houses the machine, operates it through the Julich Supercomputing Centre (JSC), builds the water cooling and green power supply, and looks into uses for the waste heat.
  • 17 January 2023: EuroHPC opens the tender, with an estimated value of EUR 273 million and a closing date of 13 February 2023. Bidders had to show experience acquiring, delivering, installing and maintaining supercomputers in similar environments.
  • 3 October 2023: EuroHPC announces the signed procurement contract with a consortium led by Eviden, with ParTec as partner. The contract was for a modular design: an NVIDIA-based Booster and a Cluster built on SiPearl’s Rhea processor.
  • December 2023: construction begins, according to the Julich inauguration release.
  • April to May 2024: the single-rack development instrument JEDI is installed and announced on 13 May 2024.
  • 18 November 2024: Julich presents JETI, a ten-rack module, as roughly 8 percent of the final Booster.
  • June 2025: the first TOP500 entry, fourth, at 793.4 PFlop/s according to The Next Platform’s reading of the list.
  • 5 September 2025: the inauguration, attended by the German Chancellor and the European Commission.
  • 17 November 2025: the 1.000 exaflop/s result.

Two budget figures circulate, and they are not the same thing. The EuroHPC tender page gives EUR 273 million as the value of the procurement. The hosting agreement and Julich’s own JUPITER page give up to EUR 500 million as the total, split EUR 250 million EuroHPC, EUR 125 million from the German federal research ministry and EUR 125 million from North Rhine-Westphalia. Heise describes the 500 million as including operating resources, the data centre and electricity over the following years. The Next Platform splits it as EUR 273 million for hardware, software and services and EUR 227 million for operations, power and cooling. That split is the Next Platform’s, and we found no official document that states the 227 million line item.

Compute time follows the money: Julich says half is managed by EuroHPC and half is awarded by the Gauss Centre for Supercomputing to German institutions.

What is actually inside

JUPITER is a modular system. The Booster is the part that was benchmarked.

Booster. Julich’s technical page lists 5,884 nodes, each with four NVIDIA GH200 superchips, so 23,536 superchips. Each superchip pairs a 72-core Grace CPU (120 GB LPDDR5X) with a Hopper GPU (96 GB HBM3). By our arithmetic that is 864 GB of memory per node and about 5.1 PB across the Booster. The TOP500 core count of 4,801,344 divides exactly by 23,536 to give 204 per superchip, which matches the way TOP500 counts GPU streaming multiprocessors alongside CPU cores. Julich states a design target of more than 70 exaflop/s of theoretical 8-bit performance. Later Julich material says “more than 40” (see below).

Packaging. The Booster is about 125 BullSequana XH3000 racks. Heise reports 192 superchips per rack and 160 kW per rack, with 36 degrees C water in and 43 degrees C out. The cooling is warm-water direct liquid cooling, and the waste heat is intended for the Julich campus heating network.

Fabric. NVIDIA Quantum-2 InfiniBand, four NDR200 (200 Gbit/s) links per Booster node, in a DragonFly+ topology. Julich’s description is 25 groups for the Booster plus 2 for the Cluster and storage, 867 high-radix switches, about 25,000 endpoints and more than 11,000 400 Gb/s global links. The Next Platform’s version has 25 cells of five XH3000 cabinets each, which agrees with 125 racks.

Cluster. This is the weakly documented part. Julich’s technical page lists 162 nodes with two SiPearl Rhea1 processors each (80 Arm Neoverse cores, 64 GB HBM2e per processor) and 582 dual-socket AMD Turin nodes (96 cores per CPU). The Next Platform in June 2025 described more than 1,300 Rhea1 nodes, and heise reports more than 2,600 planned Rhea1 chips. Those figures are the same plan, 1,300 nodes of two chips, and they are about eight times the 162 nodes on Julich’s current page. Nobody has said publicly whether the plan was scaled back, replaced by the Turin nodes, or only reordered.

Storage. ExaFLASH is 29 PB raw, about 20 PB usable, with a target of over 2 TB/s write and 3 TB/s read. Julich says it runs IBM Storage Scale on 20 IBM SSS 6000 building blocks. Behind it sit ExaSTORE, 308 PB of raw disk, and ExaTAPE at 379 PB. The Next Platform notes that only the flash was in the core procurement, and quotes 300 PB disk and 700 PB tape; we treat Julich’s figures as authoritative and note the gap.

Software. Red Hat Enterprise Linux, Slurm, Kubernetes, and ParTec’s ParaStation Modulo for running jobs across modules.

How it performs and what it does

HPL measures dense FP64 linear algebra, which is the number the exascale label rests on. JUPITER’s 1.000 exaflop/s on 17 November 2025 came about 25 months after contract signature and 2 months after inauguration. The June 2025 figure had been 793.4 PFlop/s against a 930 PFlop/s Rpeak (85.3 percent efficient, per The Next Platform). The current TOP500 page shows 1,226.28 PFlop/s Rpeak, which makes the 1.000 result 81.5 percent efficient. The Rpeak differs between readings even though the superchip count implied by the core count appears the same, and none of the sources we fetched explains why.

On power, TOP500 lists 15,793.81 kW, which works out to 63.3 GFlops/W. EuroHPC reports “more than 63” billion operations per watt for the Booster and calls it the most energy-efficient exascale system. Julich had projected about 11 MW average draw for the Booster in 2024. Those two figures measure different things (an expected average versus power during an HPL run), and the lab has not published a measured operating average that we could find. EuroHPC’s June 2026 release places the system 17th on the Green500.

The AI number is looser. The 8-bit figure is a theoretical peak, not a benchmark. Julich’s technical page says more than 70 exaflop/s, heise repeats 70, while Julich’s inauguration and TOP500 releases say more than 40. A factor-of-two gap of that kind is what sparse versus dense accounting produces, but we have not seen Julich state the basis.

Published work so far:

  • Climate: an ICON configuration won the Gordon Bell Prize for Climate Modelling at SC25. On 20,480 superchips it simulated a 1 km coupled Earth system at about 146 days of model time per 24 hours of compute (NVIDIA’s account).
  • Neuroscience: the CytoNet brain foundation model trained in under five days on 4,096 superchips using 6.5 PB of data from 21 post-mortem brains.
  • Quantum simulation: JUQCS-50, the first simulation of a 50-qubit universal quantum computer; the earlier record was 48.
  • Early access: Julich reports over 100 projects. ECMWF’s allocations total 1,064,000 node-hours across three calls.

The supply chain

  • NVIDIA: GH200 superchips (Grace and Hopper), Quantum-2 InfiniBand (the network business came from Mellanox). The domestic substitution analysis lists JUPITER as NVIDIA Grace Hopper.
  • Eviden (a unit of Atos), whose HPC line is Bull: BullSequana XH3000 and the direct liquid cooling design. TOP500 lists the manufacturer as Bull. The dataset’s news table records the French state completing its acquisition of Bull from Atos on 31 March 2026.
  • ParTec AG: system design (dynamic Modular System Architecture), the CPU module and the flash module, and the ParaStation software. ParTec has no row in our company table, so it is not linked.
  • IBM: ExaFLASH storage hardware and Storage Scale software.
  • Red Hat: RHEL 9 on the system, per TOP500 and Julich’s configuration page.
  • SiPearl: Rhea1, the European Cluster CPU. Bull’s HPC lead calls the 22 September 2026 delivery of first samples a milestone.
  • AMD: Turin CPUs in the Cluster, as listed on Julich’s technical page.
  • EuroHPC is the buyer and owner; Forschungszentrum Julich operates.

The AI-factory add-ons

On 12 March 2025 Julich announced the JUPITER AI Factory (JAIF), about EUR 55 million from EuroHPC, the federal research ministry, North Rhine-Westphalia and Hesse. It includes JARVIS, a cloud-style inference platform described as a third JUPITER module. The consortium is coordinated by JSC with RWTH Aachen’s Center for AI, two Fraunhofer institutes, hessian.AI, and associated partners WestAI and the German AI Association. Its stated purpose is to give German and European start-ups, SMEs and industry access to the machine. It adds access and services rather than a new set of processors, and we found no published hardware list for JARVIS.

What does not add up or is not public

  • The Cluster module. Node counts differ by roughly a factor of eight between 2025 and 2026 descriptions, and no Cluster HPL result has been published. The Next Platform’s June 2025 piece estimated about 5 PFlop/s for it.
  • Rhea1 timing. The Next Platform in June 2025 described a delay, and heise gave 2026 or 2027 availability. Samples arrived in September 2026. Whether the full complement ships is not stated.
  • Rpeak and the June-to-November jump. 930 PFlop/s Rpeak in one reading, 1,226.28 in another, with the same apparent chip count. Not explained.
  • Interconnect generation. Julich and TOP500 say NDR200 and Quantum-2. NVIDIA’s June 2026 article says Quantum-X800 networking. We take the lab and TOP500 as correct but note that NVIDIA’s text disagrees.
  • The budget. 273 million versus 500 million, as above. The operating-cost line is a single trade-press source.
  • JEDI’s node count. Julich’s release says 48 nodes (24 plus 24 added in May 2024). The TOP500 core count of 19,584 divides by 204 to give 96 superchips, which is 24 four-superchip nodes. The Next Platform also says 96. Julich’s quoted efficiency is 72 GFlops/W, but 4.5 PFlop/s over the 66 kW average it quotes gives about 68, so the Green500 submission must rest on a different power reading.
  • Cable length. Heise says 300 km of InfiniBand cable, ParTec says over 260 km.
  • Peak figure. Heise reports about 1.4 exaflop/s FP64, which matches no TOP500 Rpeak we found.

What this dataset says

This section describes the dataset as published on 2 October 2026. The dataset changes as rows are added and corrected.

The JUPITER row carries 1,000.00 PFlop/s Rmax, 1,226.28 PFlop/s Rpeak, 4,801,344 cores and 15,793.81 kW, rated verified, with a first operational date of 2025-09 and an announcement year of 2022. It models only the Booster, so the Cluster and the CPU question above are invisible in the dataset: the component rows cover the GH200 accelerator and Grace CPU, InfiniBand NDR, Eviden as integrator and cooling supplier, RHEL, and IBM Spectrum Scale storage. By the four roles the domestic substitution test uses (CPU, accelerator, interconnect, OS), every supplier recorded for JUPITER is headquartered in the United States, which is why we have written that Europe’s index is zero. The list-appearances table holds only two JUPITER rows, TOP500 2026-06 at rank 5 and Green500 2025-06 at rank 21, so the two fourth-place TOP500 results and the later Green500 entries are not in it; see data trust for how we treat gaps like that.

The two precursor rows make the efficiency story concrete. JEDI holds 4.50 PFlop/s at 67.31 kW, first on the Green500 in both June and November 2024. JETI holds 83.14 PFlop/s at 1,310.72 kW, 18th on the TOP500 and sixth on the Green500 in November 2024. Dividing Rmax by power from the three rows gives 66.9, 63.4 and 63.3 GFlops/W for JEDI, JETI and JUPITER, so the efficiency held roughly flat from the one-rack instrument to the full machine. See the efficiency, power and cooling and build pipeline pages for how that compares with the rest of the dataset. The dataset’s JUWELS Booster row describes that machine as the one that set the Atos-and-NVIDIA pattern that MareNostrum 5 and JUPITER later followed.

Sources

deep-diveexascaleprocurementsupply-chainpower-and-cooling