Global Performance of GPU-Optimized ICON at Kilometer Scale

The demand for local-scale weather and climate information is rapidly growing due to the green energy transition, accelerating climate change, and the increasing frequency of record-breaking extreme events. Advances in high-performance computing (HPC) architectures and targeted model development now make it possible to produce global climate information at kilometer scales. Yet, major challenges remain, including calibrating and developing such models, adapting them to heterogeneous HPC systems, and managing the massive data volumes they generate. At ETH Zurich, we are porting the ICON model to GT4Py, a domain-specific language that enables improved maintainability, portability, and computational efficiency on modern HPC platforms. We present successes and challenges of this effort, along with flagship simulations, including a global 2.5 km uncoupled ICON run covering April 2020–March 2024. This simulation, based on ICON-NWP physics and benefiting from CLM community model tuning, provides new insights into mean-state biases and the representation of mesoscale phenomena such as precipitation and wind extremes, convective organization, tropical cyclones, and equatorial waves. We further show progress toward coupled modelling, continental-scale kilometer-grid modeling, and outline our vision for future regional large-eddy simulations, highlighting key challenges and priority areas for ICON development.

Andreas Prein, ETH Zürich
Moderation: Heimo Truhetz

Part of the CLM Community Assembly 2025