Computing the Full Earth System at 1km Resolution

This invited talk, delivered by Daniel Klocke of the Max Planck Institute for Meteorology, presents the first kilometer-scale simulation of the full Earth system, work that won the ACM Gordon Bell Prize for climate modelling. The motivation is to simulate rather than parameterize: resolving convective instabilities, ocean eddies, and realistic topography while coupling the cycles of energy, water, and carbon across atmosphere, land, and ocean. Using the ICON model on Europe's two most powerful Grace Hopper systems, Jupiter at Juelich (24,000 superchips) and Alps at CSCS, the team mapped Earth system components to heterogeneous hardware: the atmosphere and tightly coupled land run on GPUs with OpenACC and CUDA graphs, while the ocean and biogeochemistry, coupled only every ten simulated minutes, run essentially for free on the Grace CPUs with OpenMP. The DaCe data-centric programming framework from ETH Zurich separates domain science from optimization, halving visible code in the dynamical core and yielding a 1.3x speedup through improved memory bandwidth. The configuration comprises nearly a trillion degrees of freedom over about 330 million one-kilometer cells with a ten-second time step, achieving 146 simulated days per day on 5,120 nodes at roughly 85 percent of Jupiter, with good weak scaling and about 4.4 times lower energy per simulated day on GPUs than CPUs. A full seasonal-cycle experiment is under way, with output stored hierarchically in HEALPix format for efficient analysis and planned public release.