15:30 〜 15:45
[AAS08-06] High-Performance Numerical Weather Prediction at ECMWF
★Invited Papers
キーワード：numerical weather prediction, kilometer-scale resolution, high-performance computing, digital twins
The skill of ECMWF's weather forecasts has improved continuously for over three decades, with each decade of research and development delivering another day in lead time. Nowadays, for example, we are able to accurately predict the tracks of tropical cyclones even more than a week before they make landfall. This achievement is thanks to a multitude of technologies allowing us to observe, describe, and compute the weather with greater and greater detail and physical fidelity.
An important factor underpinning this development has been a steady increase in compute power which has enabled an increase in model resolution, from 125 km in the 1980s to 9 km today, with a prospect for 5-km resolution forecasts in the next few years. However, running such high-resolution forecasts efficiently will require careful exploitation of the next generation of supercomputers, which will be even more massively parallel and heterogeneous. This talk will be about ECMWF's efforts in computational science to enable weather forecasts at a resolution of up to 1 km to run efficiently on these supercomputers. From a software perspective, ensuring that our model can run efficiently across the spectrum of accelerator technologies, in particular GPUs, is a huge challenge. This work is unified under our internal Hybrid2024 project, the first topic of my talk. Initial efforts in this direction have allowed us to run our model on up to 6,000 GPUs on the Summit supercomputer, and I will present some results from a physical verification of these simulations, including a discussion of how to relax the hydrostatic assumption. Ideas from the field of inexact computing have prompted us to consider alternative ways of computing within our model, and so I will present our progress and plans on exploiting low precision floating-point arithmetic and emulation of model components by machine learned algorithms. Finally, the Destination Earth project provides a unifying framework and stimulus for putting these technologies into practice, and I will finish by providing an update on the progress of our first digital twins of the Earth system.
An important factor underpinning this development has been a steady increase in compute power which has enabled an increase in model resolution, from 125 km in the 1980s to 9 km today, with a prospect for 5-km resolution forecasts in the next few years. However, running such high-resolution forecasts efficiently will require careful exploitation of the next generation of supercomputers, which will be even more massively parallel and heterogeneous. This talk will be about ECMWF's efforts in computational science to enable weather forecasts at a resolution of up to 1 km to run efficiently on these supercomputers. From a software perspective, ensuring that our model can run efficiently across the spectrum of accelerator technologies, in particular GPUs, is a huge challenge. This work is unified under our internal Hybrid2024 project, the first topic of my talk. Initial efforts in this direction have allowed us to run our model on up to 6,000 GPUs on the Summit supercomputer, and I will present some results from a physical verification of these simulations, including a discussion of how to relax the hydrostatic assumption. Ideas from the field of inexact computing have prompted us to consider alternative ways of computing within our model, and so I will present our progress and plans on exploiting low precision floating-point arithmetic and emulation of model components by machine learned algorithms. Finally, the Destination Earth project provides a unifying framework and stimulus for putting these technologies into practice, and I will finish by providing an update on the progress of our first digital twins of the Earth system.