Júpiter mapea 86 mil millones de neuronas y simula una computadora cuántica de 50 qubits

JUPITER trained a brain foundation model on 6.5 petabytes of data from 21 post-mortem brains in under five days, using 4,096 NVIDIA Grace Hopper Superchips. That model, CytoNet, maps the human brain at cellular scale, linking individual cell structures to broader functional organization across 86 billion neurons and roughly 100 trillion connections.

Neuroscientist Katrin Amunts and computer scientist Christian Schiffer at Jülich's Institute of Neuroscience and Medicine led the work. Their next step: an AI agent that integrates multimodal reasoning and language interfaces to let scientists interrogate brain data directly. "We're building an agent that can think through the experiment itself," Amunts said. JUPITER made that sentence possible.

Coupled Earth Climate at Kilometer Resolution

A novel ICON configuration running on 20,480 Grace Hopper chips on JUPITER won the Gordon Bell Prize for Climate Modelling at SC25. ICON simulates the coupled Earth system ocean, atmosphere, land, biogeochemistry, and full carbon cycle at 1-km global resolution.

Previous systems modeled pieces in isolation. ICON runs the whole system, resolving fine-scale winds, ocean eddies, and phytoplankton blooms directly from physics rather than approximations. It simulated roughly 146 days of real climate in 24 hours of compute, a world record.

6G Gets an Exascale Partner

Ericsson and Forschungszentrum Jülich announced a collaboration in March to develop AI for 5G evolution and 6G networks, with JUPITER as the training engine. Priorities include brain-inspired architectures for low-energy network operations, AI models for Ericsson's radio and core networks, and energy-efficient inference at the radio edge using neuromorphic approaches.

The collaboration leverages JSC's modular supercomputing architecture concepts. No press-release vagueness here: concrete targets with an exascale system behind them.

50-Qubit Quantum Computer Simulated on Classical Hardware

Researchers at JSC, working with the NVIDIA Application Lab, simulated a universal 50-qubit quantum computer, surpassing the previous 48-qubit record. The JUQCS-50 simulator exploits the coherent CPU-GPU memory architecture of Grace Hopper, letting data spill seamlessly from GPU to CPU memory with minimal performance loss.

That memory span held a far larger quantum state than GPU memory alone could handle, pushing past the prior record. Since today's quantum hardware can't yet outperform classical computers on useful problems, JUQCS-50 lets researchers design and stress-test algorithms at scales that future hardware will run. Kristel Michielsen, JSC director, leads JUNIQ, the quantum user facility that will make JUQCS-50 accessible.

Exascale has moved from research category into production. From neurons to atmosphere to wireless infrastructure to quantum, JUPITER turns Europe's first exascale system into a testbed for what's next, from AI-assisted neuroscience to quantum-GPU hybrid computing.

Source: At ISC, JUPITER Shows What Exascale Science Looks Like
Domain: blogs.nvidia.com