NVIDIA touts exascale wins on Europe’s JUPITER system at ISC, spanning brain maps, climate and quantum

NVIDIA is using this week’s International Supercomputing Conference (ISC) in Hamburg to spotlight what it says is real-world proof that exascale computing has moved beyond prototypes. The company’s point of reference is JUPITER, Europe’s first exascale supercomputer at Forschungszentrum Jülich, and a set of high-profile projects running on it that span brain science, Earth system simulation, AI for wireless networks and large-scale quantum computing research.

JUPITER’s compute stack is built around NVIDIA Grace Hopper Superchips and NVIDIA Quantum-X800 InfiniBand networking. In NVIDIA’s telling, the most telling throughline across four projects is that problems previously considered too large or too complex for existing hardware are now tractable at exascale, with systems supporting both science workloads and AI-driven modeling.

In neuroscience, a Jülich-led effort aims to map the brain’s microarchitecture using a foundation model called CytoNet. NVIDIA describes the research as learning from brain imaging data at cellular scale, producing a map that links individual cell structures to broader patterns of brain organization and function. The company ties the work to the Jülich Brain Atlas program, anchored at Jülich’s Institute of Neuroscience and Medicine with Helmholtz AI, partner hospitals and other Helmholtz institutions.

The model training was completed on JUPITER in under five days, using 6.5 petabytes of data from 21 post-mortem brains and running on 4,096 NVIDIA Grace Hopper Superchips. NVIDIA says the effort used CytoNet as a stepping stone toward an AI “agent” that can reason through experiments and help scientists interrogate brain data directly, including through open models and language interfaces.

NVIDIA also quoted researchers describing a shift from AI as an analysis tool to AI as an experiment-aware assistant. Katrin Amunts, director of INM-1 at Forschungszentrum Jülich, said the next step is building an AI agent for brain researchers with multimodal reasoning, language and question-answering capabilities, and it references NVIDIA Nemotron 3 120B, which NVIDIA frames as part of the open-model approach for those interfaces.

Climate modeling is the second headline application. NVIDIA says a new ICON configuration, developed by researchers including ETH Zurich, DKRZ, Jülich, the Max Planck Institute for Meteorology, NVIDIA, CSCS and the University of Hamburg, won the Gordon Bell Prize for Climate Modelling at SC25 last November. The ICON work is described as the first model to simulate a coupled Earth system at 1-kilometer resolution, including ocean, atmosphere, land, biogeochemistry and the full carbon cycle with carbon exchange among components.

NVIDIA’s description emphasizes that the breakthrough is not only finer resolution but end-to-end coupling at that scale. The company says ICON can simulate and visualize complete ecosystems, such as phytoplankton blooms and zooplankton grazing, rather than treating parts of the Earth system in isolation. It also says the approach enabled a world record in global climate simulation, with a run that converted roughly 146 days of real climate into 24 hours of compute, using 20,480 NVIDIA Grace Hopper Superchips.

On wireless and communications, NVIDIA points to a collaboration announced in March between Ericsson and Forschungszentrum Jülich to develop AI for the continued evolution of 5G and for 6G networks, using JUPITER as the compute engine for large-scale AI model training and testing. NVIDIA says the research targets brain-inspired architectures designed to handle complex network operations at lower energy costs, including AI models for Ericsson’s radio and core networks, energy-efficient AI inference at the radio edge using neuromorphic approaches, and modular supercomputing concepts drawn from JSC’s exascale work.

Finally, NVIDIA highlights progress in quantum simulation. It says researchers at the Jülich Supercomputing Centre, working with the jointly run NVIDIA Application Lab, achieved a world first by fully simulating a universal 50-qubit quantum computer, surpassing a previous 48-qubit record. The company attributes the advance to JUPITER’s tightly coupled CPU-GPU memory architecture on the GH200 Grace Hopper Superchips, which allows data to spill between GPU and CPU memory with minimal performance loss, enabling a larger coherent quantum state than GPU memory alone would allow.

A remaining caveat is what NVIDIA does not provide in its summary: the company does not detail error rates, benchmarking methodology, or how the quantum simulation results translate into real-world performance on future quantum hardware. Similarly, while NVIDIA cites a training time and dataset size for the brain foundation model, it does not provide accuracy metrics or clinical validation in the post. For investors and industry watchers, the immediate takeaway is that NVIDIA and Jülich are positioning JUPITER as a platform for increasingly production-like “exascale as workhorse” science and AI, and the next question will be how quickly these prototypes mature into repeatable workflows across more centers and use cases.