Speaker
Description
For many decades now, the European Centre for Medium-Range Weather Forecasts (ECMWF) has spearheaded developments in global numerical weather prediction. The continued growth in forecast skill over the past few decades is due in large part to increases in the grid resolution of ECMWF’s Earth-systemnmodel, the Integrated Forecasting System (IFS). This increase has gone hand-in-hand with developments innhigh-performance computing, with new generations of supercomputer permitting higher model resolutionsnand complexity. Weather forecast skill is especially sensitive to the resolution of the atmospheric component, for which resolutions are approaching the so-called “storm-resolving” level, which indicates a grid spacing ofnlower than 10 kilometres. A step change in the fidelity of global atmospheric simulations is expected as thenmodel resolution approaches this “kilometre scale”, in particular for the representation of extreme weather events.
However, recent developments in supercomputing present a barrier as we push towards these kilometre-scale simulations. The impetus for this new class of forecast system comes from the Destination Earth project, whose goals are to develop a series of Earth-system digital twins to aid in the prediction and mitigation of extreme weather events under a changing climate. In order to run this new class of Earth-system simulation efficiently, one must make effective use of accelerators, namely GPUs, and large communication networks.
This talk will give an overview of activities at ECMWF towards the goal of running kilometre-scale Earth-system simulations on pre-exascale and exascale supercomputers. The talk will present lessons learned from earlier experiments on supercomputers such as Summit. I will concentrate in particular on the spectral transform library ecTrans which the IFS atmospheric component crucially depends on, and which neatly contains several key computation and communication paradigms. I will also explore the opportunities of the new breed of data-driven models, which are led by ECMWF’s AIFS machine learning model. These models rival traditional “physics-based” models such as the IFS, and are extremely cheap at inference time. The training of these models is a high-performance computing problem in its own right.
Sam Hatfield is a Computational Scientist at the European Centre for Medium-Range Weather Forecasts (ECMWF). He is closely involved in efforts to port the Integrated Forecasting System (IFS), a numerical weather prediction code with a long history, to modern GPU-equipped supercomputers. This will enable ECMWF to perform kilometer-scale global Earth system simulations to better capture extreme weather events, a key goal of the Destination Earth initiative. Sam focuses his efforts in particular on the spectral transform kernel, fulfilled by ECMWF’s numerical library ecTrans, which can dominate the overall wall time for the IFS.
