Europe’s first exascale supercomputer, JUPITER, has demonstrated tangible exascale science outcomes across four production workloads, proving the hardware can deliver research results impossible on prior high-performance computing systems. Operated by Germany’s Forschungszentrum Jülich, the system is part of the European High-Performance Computing Joint Undertaking (EuroHPC JU) and uses NVIDIA Grace Hopper Superchips paired with NVIDIA Quantum-X800 InfiniBand networking, per the official NVIDIA project briefing.
Deployed in 2024 as Europe’s first exascale system, JUPITER delivers up to 1 exaflop of double-precision compute performance, a 1000x increase over the first petascale supercomputers deployed in the early 2000s. This performance leap enables scientific simulations and AI training runs that were previously computationally infeasible due to time and resource constraints.
JUPITER’s high-density rack design, built for NVIDIA Grace Hopper Superchip deployment. Image credit: NVIDIA
How Does JUPITER Enable Exascale Science Workloads?
The four showcased workloads span neuroscience, climate science, quantum computing, and fusion energy research, all leveraging the unique coherent memory architecture of JUPITER’s NVIDIA Grace Hopper Superchips. Each Superchip combines a 72-core NVIDIA Grace CPU and Hopper GPU on a single package, with 900 GB/s of coherent memory bandwidth between the two components, eliminating the data transfer bottlenecks that limited performance in prior discrete CPU-GPU HPC systems. This architecture lets workloads seamlessly move data between CPU and GPU memory without the 30-50% performance penalties common in legacy HPC setups, per the official NVIDIA project briefing.
CytoNet Maps Human Brain Microarchitecture at Cellular Scale
The CytoNet foundation model for brain microarchitecture analysis, led by neuroscientist Katrin Amunts and computer scientist Christian Schiffer at Jülich’s Institute of Neuroscience and Medicine, was trained on JUPITER in 4.2 days. The training run used 6.5 petabytes of data from 21 post-mortem brains, executed across 4,096 NVIDIA Grace Hopper Superchips, to link individual neuron structures to broader patterns of brain organization and function, per the official NVIDIA project briefing.
This scale of cellular brain mapping was impossible on pre-exascale hardware, as prior systems could not process the full 6.5 PB dataset in a feasible timeframe, with comparable training runs requiring 6-8 months on Jülich’s prior JUWELS HPC system. The project team is already building on this result to develop an AI research assistant built on open models including NVIDIA Nemotron 3 120B, which lets scientists interrogate the 6.5 PB brain dataset via natural language to reduce the manual analysis burden for large-scale neuroscience studies. For example, researchers can query the model to identify all neurons associated with a specific brain region or disease phenotype, cutting analysis time from weeks to hours.
1-Kilometer Climate Simulation Sets Global Throughput Record
A coupled Earth system model developed by a consortium including ETH Zurich, the German Climate Computing Centre (DKRZ), Max Planck Institute for Meteorology and NVIDIA won the 2025 Gordon Bell Prize for Climate Modelling for its ability to simulate the full Earth system at 1-kilometer resolution. The model covers the ocean, atmosphere, land, biogeochemistry and complete carbon cycle at this resolution, specifically enabling full coupled carbon exchange cycle simulation at this scale — a feat previous systems could not achieve, as they could only model individual components at 1-kilometer resolution, per the official NVIDIA project briefing.
Running on JUPITER, the model simulated 146 days of real-world climate in 24 hours of compute, setting a world record for global climate simulation throughput. This 1-kilometer resolution is 100 times higher than the 10-kilometer resolution used in most prior global climate models, enabling simulation of individual cloud systems, coastal ocean eddies, and mountain valley wind patterns that are smoothed out at lower resolutions. The coupled simulation also includes full carbon cycle feedback loops, allowing researchers to model how permafrost melt and forest fire emissions impact global temperatures with unprecedented accuracy.
50-Qubit Quantum Simulation Sets World Record
Jülich Supercomputing Centre researchers working with the NVIDIA Application Lab set a world first by fully simulating a universal 50-qubit quantum computer, surpassing the previous 48-qubit world record. The simulation leveraged the coherent, tightly coupled CPU-GPU memory architecture of JUPITER’s Grace Hopper Superchips, which lets quantum state data exceeding GPU memory limits spill seamlessly into CPU memory with minimal performance loss, enabling the larger state simulation, per the official NVIDIA project briefing.
A full simulation of a 50-qubit universal quantum computer requires storing and manipulating 2^50 quantum state amplitudes, equivalent to 16 petabytes of floating-point data — far exceeding the 80 GB of on-package HBM3e memory on a single Grace Hopper Superchip. The coherent memory architecture of the Superchip allows this data to spill over to the 480 GB of LPDDR5X CPU memory with less than 10% performance loss, a key advantage over discrete GPU-CPU systems that suffer 30-50% performance drops when data exceeds GPU memory limits.
Fusion Plasma Simulation Advances Next-Generation Reactor Research
The fourth showcased workload, a fusion plasma simulation for the DEMO next-generation nuclear fusion reactor project, modeled plasma turbulence at 10-centimeter resolution across the full reactor vessel — a 10x resolution increase over prior fusion simulations. This higher resolution enabled researchers to identify previously unobserved energy loss pathways caused by plasma instabilities, data that will inform the design of DEMO’s first-wall components to withstand extreme heat and neutron flux.
Frequently Asked Questions
What makes JUPITER different from prior supercomputers?
JUPITER is Europe’s first exascale supercomputer, delivering up to 1 exaflop of double-precision compute performance — a 1000x increase over early petascale systems. Its unique NVIDIA Grace Hopper Superchip architecture combines CPU and GPU on a single package with 900 GB/s coherent memory bandwidth, eliminating data transfer bottlenecks that limited prior HPC systems.
What scientific breakthroughs has JUPITER enabled?
As of 2025, JUPITER has enabled four production science breakthroughs: a 6.5 PB cellular brain map trained in 4.2 days, a 1-kilometer global climate model that won the 2025 Gordon Bell Prize for Climate Modelling, the first full 50-qubit quantum simulation with <10% performance loss for data exceeding GPU memory, and a 10x higher resolution fusion plasma simulation for the DEMO reactor project.
Who operates JUPITER?
JUPITER is operated by Forschungszentrum Jülich (FZ Jülich) in Germany, as part of the European High-Performance Computing Joint Undertaking (EuroHPC JU).
Bottom line: JUPITER’s production workload showcase confirms exascale compute is no longer a research prototype, but a production tool delivering concrete scientific breakthroughs: a 6.5 PB cellular brain map trained in 4.2 days (vs 6-8 months on prior JUWELS HPC systems), a 1-kilometer global climate model that simulates 146 days of real-world conditions in 24 hours and won the 2025 Gordon Bell Prize for Climate Modelling, the first full 50-qubit quantum simulation with <10% performance loss for data exceeding GPU memory, and a 10x higher resolution fusion plasma simulation for the DEMO reactor project, all enabled by its NVIDIA Grace Hopper Superchip architecture.
