ACM Gordon Bell Prize finalist: BEAST model harnesses LUMI’s power

A research project harnessing the computing power of the LUMI supercomputer has been shortlisted for the prestigious ACM Gordon Bell Prize for Climate Modelling 2026, one of the world’s most renowned awards in high-performance computing. The project was submitted by an international research team led by Karlsruhe Institute of Technology (KIT), bringing together researchers and experts from Jülich Supercomputing Centre (JSC), NVIDIA and CSC – IT Center for Science.

The nominated work introduces BEAST (Bayesian Exascale Atmospheric Swin Transformer), an AI model designed for high-resolution global atmospheric forecasting that can also quantify the uncertainty of its predictions. This capability is particularly important for forecasting rare and high-impact phenomena such as tropical cyclones, heatwaves and cold spells. The model was trained on 40 years of atmospheric data and operates at a global resolution of approximately 30 kilometres.

A central achievement of the project is a novel four-dimensional parallelisation strategy that enables large Bayesian AI models to be trained efficiently on modern supercomputers. Bayesian models provide valuable information about forecast uncertainty but require significantly more computational resources than conventional AI approaches. The new method makes it possible to scale these calculations across thousands of GPUs while maintaining high efficiency.

The team’s work demonstrated outstanding performance on two of Europe’s leading supercomputing systems. On the JUPITER supercomputer, the researchers achieved 3.96 exaflops of mixed-precision performance using 20,480 NVIDIA GH200 GPUs with a 2.4-billion-parameter model. On the LUMI supercomputer, training reached 463.23 petaflops using 4,096 AMD MI250X GPUs. These results highlight the ability of the new approach to fully exploit next-generation HPC infrastructure.

Specialists from CSC played a key role in adapting and scaling the application on LUMI, where the team achieved 92% strong-scaling efficiency across 1,024 nodes. The work demonstrates how LUMI can support cutting-edge AI research at unprecedented scale while advancing weather and climate modelling.

In addition to its computational performance, BEAST delivers forecast quality comparable to leading probabilistic AI weather models and established numerical weather prediction systems. At the same time, it can generate large forecast ensembles several times faster than comparable AI-based forecasting approaches, enabling more efficient assessment of uncertainty and extreme weather risks.

The ACM Gordon Bell Prize for Climate Modelling recognises innovative uses of high-performance computing that advance climate science and help address global climate challenges. The nomination highlights the role of European leadership-class supercomputers such as LUMI and JUPITER in advancing next-generation AI methods for weather and climate research.

– It’s been great to be involved in cutting-edge AI research using LUMI, running large-scale training experiments and working together with an excellent team. And of course, we’re very excited about the Gordon Bell nomination, says Anni Moisala, one of the experts involved in the project from CSC.

The Gordon Bell Prize winners will be announced at the SC26 conference in Chicago in November 2026.

Preprint available: 4D Parallelism Unlocks Exascale Bayesian Neural Networks for High-Fidelity Atmospheric Modeling