JUPITER, Europe’s first exascale supercomputer at Germany’s Forschungszentrum Jülich, runs on NVIDIA Grace Hopper Superchips and NVIDIA Quantum-X800 InfiniBand networking — and it’s had a busy 12 months.
Because the worldwide supercomputing neighborhood gathers at ISC in Hamburg this week, 4 tasks working on JUPITER level to what exascale computing can really do: map the human mind at mobile scale, simulate your entire Earth’s local weather at 1-kilometer decision, construct AI programs for the subsequent era of wi-fi networks and simulate a common 50-qubit quantum pc.
“With JUPITER, Europe doesn’t simply be a part of the exascale period — it leads it, throughout the widest vary of science and AI of any system worldwide,” mentioned Thomas Lippert, director of the Jülich Supercomputing Centre and professor at Goethe College Frankfurt.
4 tasks, detailed beneath, share a throughline: scientific issues that have been out of attain on earlier {hardware} are actually tractable at exascale.
A Basis Mannequin for Mapping the Mind
The Jülich Mind Atlas undertaking — anchored at Jülich’s Institute of Neuroscience and Medication with companion Helmholtz AI, companion hospital and different Helmholtz establishments — has produced CytoNet, a basis mannequin for mind microarchitecture evaluation.
The complexity of the human mind is astonishing. With 86 billion neurons and about 100 trillion connections between them, understanding mind operate at single neuron decision has been out of attain, till now.
The analysis is led by neuroscientist Katrin Amunts and pc scientist Christian Schiffer at INM-1, Jülich’s Institute of Neuroscience and Medication. The mannequin learns from mind imaging information at mobile scale, constructing a map that hyperlinks particular person cell constructions to broader patterns of mind group and performance.
Coaching ran on JUPITER in underneath 5 days, utilizing 6.5 petabytes of information from 21 autopsy brains on 4,096 NVIDIA Grace Hopper Superchips. A paper describing the work is obtainable on arXiv.
“For the primary time, we’re not simply utilizing AI to research the mind — we’re constructing an agent that may suppose by the experiment itself,” mentioned Katrin Amunts, director of INM-1 at Forschungszentrum Jülich and professor of mind analysis at Heinrich Heine College Düsseldorf. “That adjustments what neuroscience will likely be, and JUPITER is what makes that sentence potential to say as we speak.”
That agent is the workforce’s subsequent step: constructing an AI agent for mind researchers — integrating multimodal reasoning, language interfaces and Q&A capabilities utilizing open fashions, together with NVIDIA Nemotron 3 120B, working towards AI assistants that may assist scientists interrogate mind information instantly.
Local weather at Kilometer Decision
A novel ICON configuration — developed by researchers on the ETH Zurich, German Local weather Computing Centre (DKRZ), Jülich Supercomputing Centre (JSC), Max Planck Institute for Meteorology, NVIDIA, Swiss Nationwide Supercomputing Centre (CSCS) and the College of Hamburg — received the Gordon Bell Prize for Local weather Modelling at SC25 final November.
The breakthrough isn’t decision alone. ICON is the primary mannequin to simulate a coupled Earth system at 1-kilometer decision, with ocean, ambiance and land, biogeochemistry and the complete carbon cycle, with carbon exchanged, between all parts. It might probably simulate and visualize full ecosystems, resembling phytoplankton blooms and zooplankton grazing. Earlier programs might mannequin items of this; ICON runs all of it. This enables a way more exact and full simulation of the Earth — observable at that stage of element for the primary time.
Working on 20,480 NVIDIA Grace Hopper Superchips on JUPITER, the mannequin simulated roughly 146 days of actual local weather into 24 hours of compute, setting a world file in world local weather simulation. NVIDIA’s involvement within the ICON neighborhood spans greater than a decade.
“Our simulations resolve the fine-scale winds, ocean eddies and upper-ocean mixing that form marine ecosystems and regulate the ocean’s uptake of carbon,” mentioned Daniel Klocke, computational infrastructure and mannequin improvement group chief on the Max Planck Institute for Meteorology. “At a world decision of simply 1 kilometer, many of those interactions emerge instantly from the legal guidelines of physics slightly than being approximated. This offers us an unprecedented view of how the ambiance, ocean and biosphere work collectively, serving to us perceive the processes driving our altering local weather.”
6G Will get an Exascale Companion
In March, Ericsson and Forschungszentrum Jülich introduced a collaboration to develop AI for the continued evolution of 5G and for 6G networks — with JUPITER because the compute engine for large-scale AI mannequin coaching and testing.
The collaboration targets brain-inspired architectures designed to deal with complicated community operations at far decrease power prices.
Analysis priorities embrace AI fashions for Ericsson’s radio and core networks, energy-efficient AI inference on the radio edge utilizing neuromorphic approaches, and modular supercomputing structure ideas drawn from JSC’s exascale work.
Breaking Quantum Data
Researchers on the Jülich Supercomputing Centre (JSC), working with the collectively run NVIDIA Utility Lab, additionally set a world first by totally simulating a common 50-qubit quantum pc, surpassing the earlier 48-qubit file.
The simulation was made potential by drawing on the coherent, tightly coupled CPU-GPU reminiscence structure of JUPITER’s NVIDIA GH200 Grace Hopper Superchips, which lets information exceeding GPU limits spill seamlessly into CPU reminiscence with minimal efficiency loss — permitting JUPITER to carry a far larger quantum state than GPU reminiscence alone, which is what pushed the simulation previous the earlier 48-qubit file.
For now, that form of simulation is essentially the most highly effective instrument quantum analysis has: as we speak’s quantum {hardware} can’t but outperform classical computer systems on helpful issues, so simulating quantum machines on the largest potential scale is how researchers design and stress-test the algorithms that future {hardware} will run.
This highly effective quantum simulator, JUQCS-50, will likely be accessible to discover the efficiency of quantum algorithm designs inside JUNIQ, the quantum pc consumer facility at JSC, led by Kristel Michielsen, director of JSC and professor on the College of Cologne. JUQCS-50 turns Europe’s first exascale system into a robust testbed for tomorrow’s quantum-GPU supercomputers.
Exascale’s Influence
The vary of science working on JUPITER — from neurons to ambiance to wi-fi infrastructure to quantum — makes a case that exascale computing has moved from a analysis class into manufacturing.
The outcomes are a proof level for the Grace Hopper platform on the frontier of science.
Study extra about NVIDIA AI for science.
Source link
#ISC #JUPITER #Shows #Exascale #Science
