Abstract
Predicting the long-term, localized effects of climate change requires a transition from current 10-kilometer global models to kilometer-scale resolutions. This paper provides a rigorous analysis of the ICON (Icosahedral Nonhydrostatic) model to determine the computational, memory, and storage requirements for next-generation climate science. We find that a 30-year simulation at 1.23 km resolution requires approximately 11 Zettaflops of computing power and generates 14.4 Petabytes of data, highlighting a significant gap between current capabilities and scientific needs.
Introduction: Bridging the Resolution Gap
Climate modeling currently faces a "resolution gap" that limits our ability to predict localized phenomena like storm surges or cloud-level physics. To capture these processes from first principles, models must operate at a 1-kilometer scale. However, doubling horizontal resolution increases computational demands by a factor of 8 to 16. Current models operate at resolutions where large thunderstorms are barely visible, and smaller clouds are essentially "guessed at" using approximations known as parameterizations.
Resolving clouds and global circulation requires moving beyond these approximations to direct simulation. This necessitates a "bottom-up" understanding of the computational costs associated with the most advanced modeling frameworks, such as the ICON model developed by ETH Zurich and MeteoSwiss.
The ICON Framework and Methodology
The ICON model utilizes an icosahedral grid to tile the Earth. This approach, based on the recursive bisection of a 20-sided solid, ensures a relatively uniform distribution of points across the sphere, avoiding the "pole problem" found in traditional latitude-longitude grids.
The researchers dissected the ICON model—comprising over 630,000 lines of Fortran code—into its core components: the dynamical core (fluid motion), tracer transport (humidity and aerosol movement), and physical parameterizations. By instrumenting the code with hardware counters using tools like LibLSB, they developed mathematical models to project requirements for resolutions that exceed current supercomputing limits, specifically focusing on grids ranging from R2B3 to R2B11.
Key Insights: The Computational Mountain
The analysis reveals several critical bottlenecks that must be addressed to reach the "holy grail" of 1km resolution:
- The 11 Zflop Milestone: A 30-year climate run at 1.23 km resolution requires ~11 Zettaflops ($10^{21}$ operations). For context, if an exascale system operated at 1% efficiency, it would take 13 days of the entire system's time to finish just one simulation trajectory.
- The Communication Wall: As resolution increases, the dynamical core's importance grows faster than other components. At 1.23 km resolution, the total communication volume for a 30-year run is approximately 3.66 PiB, highlighting the need for optimized MPI communication patterns.
- Storage and I/O Sustainability: Generating 14.4 PiB of data per simulation is unsustainable for current storage architectures. The study suggests that future workflows must either stream data directly into machine learning models or employ aggressive lossy compression techniques to maintain an average I/O bandwidth of 12.8 GiB/s.
Conclusion and Future Directions
To achieve kilometer-scale climate modeling, we need 3-4 orders of magnitude more computing power than is currently available. The authors argue that "ML inside" (replacing expensive physics modules with neural networks) or "ML on top" (using AI for post-processing) is the only practical path forward. This paper provides the essential roadmap for building the specialized supercomputers that will generate the high-resolution training data required for these future AI-driven climate models.
Summary updated