Germany is building the Jupiter exascale supercomputer as part of Europe's strategic AI and HPC infrastructure independence initiative.

Jupiter, the long-awaited exascale-class supercomputer at Forschungszentrum Jülich facility in Germany, is the first exascale system to be completed under the EuroHPC Joint Undertaking of the European Union. While not utilizing custom CPUs and XPUs created by European companies as originally hoped, Jupiter is essentially an Nvidia machine from top to middle, with bottom storage still to be acquired. However, the Universal Cluster module will be based on the "Rhea1" Arm server CPU created by SiPearl, representing a step toward independence for European HPC. The system is built by Eviden, the HPC division of Atos, and ParTec, a German HPC system designer and installer.

Jupiter is a hybrid supercomputer following the design of its predecessor, "Jewels," first deployed in 2018 and upgraded over the years. Like Jewels, which paired Intel "Skylake" Xeon SP processors with 100 Gb/sec EDR InfiniBand in a BullSequana X1000 system and later added a Booster Module with AMD "Rome" Epyc CPUs and Nvidia "Ampere" GPUs with 200 Gb/sec HDR InfiniBand, Jupiter combines modular CPU and GPU compute blocks with storage and acceleration components.

The Universal Cluster will include more than 1,300 CPU-only nodes, each based on a pair of Rhea1 chips with 80 cores each using the "Zeus" Neoverse V1 design, the same cores in Amazon Web Services' "Graviton3" Arm chip with pairs of 256-bit SVE vector engines. Each Rhea1 features 64 GB of HBM memory. The Rhea1 chip was delayed in June 2024 and is expected sometime later in 2025 for FZJ. A variant of the SiPearl Arm CPU will also be employed in a second European exascale system, "Alice Recoque," set to be hosted in France and built by Eviden, with a budget of €542 million ($580.2 million) for the system, facility, power, and cooling. The Universal Cluster is expected to achieve 5 petaflops of FP64 performance on the HPL benchmark, approximately 7 petaflops at peak theoretical performance.

The Jupiter GPU Booster module, which placed fourth on the June 2025 Top500 rankings in the High Performance LINPACK benchmark, is based on a unique four-way clustering of Nvidia "Grace" G100 Arm server CPUs using four "Hopper" H200 GPUs as NUMA node controllers, linking eight compute engines in total. Each Jupiter node contains two compute sleds, each with a pair of Grace-Hopper superchips. The H200 GPUs have 96 GB of HBM3 memory each with 4 TB/sec of bandwidth, while NVLink 4 ports cross-linking the four H200s provide 300 GB/sec of bandwidth (150 GB/sec in each direction). Each Hopper GPU communicates with its associated Grace CPU at 600 GB/sec (300 GB/sec per direction) and with the other three CPUs in the quad at 100 GB/sec (50 GB/sec). Each CPU has a PCI-Express 5.0 port reaching a 200 GB/sec ConnectX-6 SmartNIC, and each Jupiter node includes a pair of dual-port 400 Gb/sec ConnectX-7 NDR InfiniBand cards, with one port for each superchip.

Both the Universal Cluster and GPU Booster modules use the BullSequana XH3000 system design by Eviden, with ParTec serving as prime contractor, adding installation and services to ensure German company involvement in the project.