Researchers Achieve First 1.25-Km Global Earth Simulation, Hitting 145.7 Days Per Day
Updated
Updated · Quantum Zeitgeist · Sep 3
Researchers Achieve First 1.25-Km Global Earth Simulation, Hitting 145.7 Days Per Day
1 articles · Updated · Quantum Zeitgeist · Sep 3
Summary
A 1.25-kilometer global Earth system run on JUPITER became the first of its kind, with the ICON model reaching 145.7 simulated days per day—above the roughly 100-day threshold seen as necessary for kilometer-scale science.
GPU acceleration drove the result: researchers adapted ICON’s Fortran code with OpenACC, kept data resident on NVIDIA Grace Hopper GPUs during time-stepping, and split work so GPUs handled atmosphere and land while CPUs ran ocean and biogeochemistry.
That approach still exposes a major constraint. OpenACC support is strongest on NVIDIA’s nvhpc stack, leaving ICON poorly portable to AMD GPUs and Intel-based exascale systems such as Aurora and deepening vendor lock-in.
Projects including WarmWorld and EXCLAIM are now pushing a broader rewrite using abstraction layers, Python DSLs and Kokkos-like C++ models, following SCREAM’s example of running across AMD and NVIDIA hardware without code changes.
The shift matters because 1.25-km Earth models can involve nearly 1 trillion degrees of freedom and at least 8 tebibytes of memory, making performance portability central to future climate and weather prediction.
Could Europe's record-breaking 1.25-km climate simulation on the JUPITER supercomputer accidentally introduce artificial global warming biases through complex parameter tuning?
As climate models reach unprecedented 1.25-kilometer resolution, are we solving weather prediction or just creating an unmaintainable software nightmare?
Will abandoning NVIDIA's proprietary code for Python-based models sacrifice the groundbreaking speed that made exascale climate prediction possible?