September 29, 2025 – by Santina Russo
For the first time, scientists have performed global coupled climate simulations at a resolution of just 1.25 kilometres, fine enough to capture local phenomena such as rainstorms anywhere on Earth. More than that: on supercomputers like CSCS’s “Alps”, these high-resolution simulations can be run at decadal time scales, meaning that “simulations covering several decades are feasible, which is a quantum leap at this resolution,” said Torsten Hoefler, Chief Architect for AI and ML at CSCS. “No climate simulations have ever come close to this efficiency, fidelity, and accuracy before.”
This breakthrough, achieved by a group of researchers including Torsten Hoefler at CSCS and Alexandru Calotoiu from the Scalable Parallel Computing Lab at ETH Zürich, pushes past long-standing barriers in climate modelling. Until now, researchers had to choose between regional models that resolve small-scale weather events but only over limited areas, or global models that capture Earth’s interconnected systems but at a much coarser resolution of about 50 kilometres, which is too imprecise to represent local features. Hoefler compares the difference to the clarity of a photograph: a low-resolution image with 100×100 pixels is blurred, while a megapixel picture reveals fine details.
For this achievement, their project was selected as a finalist for the prestigious Gordon Bell Prize for Climate Modelling.
A power boost for ICON
The project team includes scientists from the Max Planck Institute for Meteorology and the German Climate Computing Center (DKRZ), both in Hamburg, and collaborators from NVIDIA. Together, the experts optimized the implementation of the ICON weather and climate prediction model on Europe’s newest and most ambitious computing systems including the CSCS “Alps” supercomputer.
The climate modelling experts developed a new ICON configuration that allowed them to include all relevant Earth system processes: the atmosphere, the oceans and their biogeochemistry, and the land areas with their different vegetation and characteristics—from desert to forests to ice sheets. “We have to model all of these things to realistically capture their impact on the carbon cycle,” explained Hoefler.
His group’s contribution to the project was applying their Data Centric Parallel Programming concept to boost the efficiency of the most computationally demanding part of the model code, the so-called dynamical core, where fundamental physical equations are used to represent the processes in the atmosphere. “Since this dynamical core consumes most of the model’s computing power, improving its efficiency has the greatest overall effect,” Hoefler said. “The resulting gain in efficiency allowed us to run the most comprehensive and highest-resolution Earth system simulations at global scale to date.”
Alps setting a high bar
In April 2025, the researchers used 8,000 of the “Alps” supercomputer’s NVIDIA Grace Hopper 200 superchips for a power run and achieved a speed of 82.5 simulated days per day of computation.
What does being a finalist mean?
The Gordon Bell Prize for Climate modelling is awarded each year to recognize outstanding achievement in high-performance computing for climate research and its impact on society by contributing to solutions for the global climate crisis. “We are very proud that Swiss researchers are among the finalists of such a prestigious prize, with their submission that includes simulations on the ‘Alps’ supercomputer,” said CSCS Director Thomas Schulthess. “This shows that we are achieving our goal of supporting Swiss research and fostering its impact on society, both here and worldwide.”
In a subsequent run in September, the team accelerated the simulations even further to 91.8 simulated days per day. “This means that high-fidelity coupled simulations covering 30 years would take us less than four months on ‘Alps’, which makes simulations capturing local effects of climate change feasible for the first time,” said Hoefler. “In terms of computing power and efficiency, the bar that we set with ‘Alps’ is extremely high.”
“
No climate simulations have ever come close to such efficiency, fidelity, and accuracy before.
”
Automating the tailoring to GPUs
This was no easy task. “Climate simulations are among the most complex simulations of all,” said Hoefler. Since the Earth’s climate is a chaotic system, the smallest change in the input can produce arbitrarily large changes in output. Scientists therefore need to run not just one, but many simulations to understand it. And these simulations need to consider very different scales—from millimetre-scale, which is covered heuristically, to thousands of kilometres, and from fractions of seconds to decades in time.
All of this is captured by the ICON model, a massive piece of software consisting of roughly one million lines of FORTRAN code. But its complexity is elegantly addressed by Hoefler and his team’s approach: their Data Centric Parallel Programming framework DaCe optimizes software components to run on specific hardware without altering the original code. The principle behind it is an automated analysis of the code to improve data movement and find items that can be parallelized. Simply put: If some statements in a code are completely independent of each other, they can be reordered for more efficient data movement or executed in parallel to boost performance. Using this approach, the team analysed and transformed the extensive dynamical core code to map it to the Grace Hopper superchips’ combined GPU-CPU design of the “Alps” supercomputer, with data movement optimized and parallelism exploited as much as possible.
Larger than AI models
“This data-centric automated approach saves hundreds of work hours that previously had to be spent manually optimizing the code for specific hardware,” said Alexandru Calotoiu, Senior Scientist in Torsten Hoefler’s group, who was mainly responsible for applying the approach to ICON. The team’s tests also suggested that their automated process delivered a 20 percent higher efficiency boost than conventional manual optimisation. “Depending on the level of optimisation, our data-centric approach reduces manual coding effort by one to two orders of magnitude” Calotoiu added.
To illustrate the scale of the challenge overcome: Modelling of the atmosphere included 347 billion parameters to capture all climate-relevant variables on a 10-second time scale. In total, the ICON simulations included 780 billion parameters, which is larger than the biggest public AI models and pushing the boundaries of what is feasible.
Merely doubling the resolution of such a simulation, for example from 10 to 5 kilometres, increases the computational requirements tenfold. Thus, moving from a resolution of 50 kilometres—the previous typical resolution of global models—to the 1.25 kilometres achieved by the project team, demanded a performance improvement by roughly 100,000-fold (10-fold for each doubling in resolution).
The team is now working to expand this scope to cover more of the ICON FORTRAN code, aiming to map all one million lines to the superchips’ GPUs, Hoefler said. “This will bring another leap in simulation efficiency and speed.”