JUPITER, Europe's first exascale supercomputer (capable of more than 1 quintillion operations per second) installed at Germany's Forschungszentrum Jülich, is producing results across fields ranging from neuroscience to climate, wireless communications and quantum computing[1]. Powered by NVIDIA Grace Hopper Superchips, the system presented four research projects at the ISC conference in Hamburg, Germany, showing that problems once out of reach for earlier machines can now actually be solved.
Computing at a Scale That Was Previously Impossible
JUPITER is built on NVIDIA Grace Hopper Superchips (chips that integrate a CPU and a GPU into a single package) and Quantum-X800 InfiniBand (a high-speed data network)[1]. In total, the system bundles roughly 24,000 GH200 Grace Hopper Superchips and is regarded as Europe's fastest supercomputer[2].
Thomas Lippert, director of the Jülich Supercomputing Centre, said, "With JUPITER, Europe doesn't just join the exascale era. It leads it, across the widest range of science and AI of any system worldwide"[1]. The four projects unveiled this time share a common thread: problems that earlier hardware could not handle have become tractable at exascale.
Mapping the Brain at the Cellular Level, with Training in Under Five Days
The first is the Jülich Brain Atlas project, which maps the human brain at the cellular level. The research team developed CytoNet, a foundation model for analyzing the microarchitecture of the brain[1].
The human brain contains about 86 billion neurons, with roughly 100 trillion connections among them. Because of this complexity, understanding brain function at the resolution of individual neurons had long been impossible[1]. Training CytoNet used 6.5 petabytes of imaging data from 21 post-mortem brains and 4,096 Grace Hopper Superchips, and took under five days[1].
Katrin Amunts, director of INM-1, said, "For the first time, we're not just using AI to analyze the brain. We're building an agent that can think through the experiment itself"[1]. The team's next step is to build an AI agent for brain researchers using open models such as NVIDIA Nemotron 3 120B.
Reproducing the Entire Earth at One-Kilometer Resolution
The second is climate simulation. The ICON climate model, jointly developed by ETH Zurich, the Max Planck Institute for Meteorology, NVIDIA and others, won the Gordon Bell Prize for Climate Modelling at SC25 last November[1].
What makes ICON groundbreaking is not resolution alone. It is the first model to reproduce a coupled Earth system, including ocean, atmosphere and land plus biogeochemistry and the full carbon cycle, at a resolution of 1 kilometer[1]. Running on 20,480 Grace Hopper Superchips, it simulated roughly 146 days of real climate in 24 hours of compute, setting a world record in global climate simulation.
Daniel Klocke of the Max Planck Institute for Meteorology explained, "At a global resolution of just 1 kilometer, many of these interactions emerge directly from the laws of physics rather than being approximated," adding that this provides an unprecedented view of how the atmosphere, ocean and biosphere work together[1].
Research on 6G and a Collaboration with Ericsson
The third is next-generation communications. In March, Ericsson and Forschungszentrum Jülich announced a collaboration to develop AI for the evolution of 5G and for 6G networks, positioning JUPITER as the compute engine for large-scale AI model training and testing[1]. Using brain-inspired architectures, the effort aims to handle complex network operations at lower energy costs. Research priorities include AI models for radio and core networks, and energy-efficient AI inference near the radio edge.
Fully Reproducing a 50-Qubit Quantum Computer
The fourth is quantum computer simulation. Researchers at JSC and the NVIDIA Application Lab achieved a world first by fully simulating a universal 50-qubit quantum computer, surpassing the previous 48-qubit record[1].
This was made possible by the tightly coupled CPU-GPU memory architecture of JUPITER's GH200 Grace Hopper Superchips. Because data exceeding GPU memory capacity can spill into CPU memory with minimal performance loss, the system can hold a far larger quantum state than a GPU alone[1]. Today's quantum hardware cannot yet outperform classical computers on useful problems, so simulating quantum machines at the largest possible scale is a powerful way to design and stress-test the algorithms that future hardware will run. The quantum simulator, JUQCS-50, will be made available within JUNIQ, JSC's quantum computing user facility[1].
Summary
From neurons to the atmosphere, wireless infrastructure and quantum, the breadth of research running on JUPITER shows that exascale computing has moved from a research category into production[1]. NVIDIA's Grace Hopper platform is beginning to deliver concrete results at the frontier of science.
出典:https://blogs.nvidia.com/blog/jupiter-exascale-supercomputing-science/
