Adapting to a growing range of high-performance computing (HPC) architectures is contributing to a transformation in our capabilities to model planet Earth.
To achieve faster computer simulations of the Earth system with a lower energy footprint, we need our forecasting codes to be optimally mapped to heterogenous HPC architectures – something that we can only achieve through our collaboration with hardware architecture experts, software experts and domain specialists.
Computer accelerators, such as graphics processing units (GPUs), offer increased computer power coupled with greater energy efficiency. Today, supercomputers are combining central processing units (CPUs) with GPUs and other types of accelerators, in a hybrid architecture to get the benefit of a range of technologies, and the promise of exascale performance.
However, it is an enormous task to optimise our whole Integrated Forecasting System (IFS) to get the most from accelerators, while ensuring it is sufficiently flexible to cope with the diversity of current and emerging technology and, at the same time, continue scientific developments.
Harnessing developments in HPC is vital to make operational kilometre-scale weather forecasts a reality, to assimilate a growing range of observations and to model the Earth system in more detail. This underpins ECMWF’s key goal of more accurate predictions of high-impact weather up to two weeks ahead, as well as wider collaborative ambitions in environmental monitoring and prediction focused in the EU Destination Earth initiative (DestinE).
Building on work within ECMWF’s Scalability Programme, the Hybrid2024 project is preparing our forecasting system for the future. Exploring a variety of approaches, core model components are being restructured to ultimately achieve a full accelerator-enabled multi-architecture IFS. The computationally expensive spectral transform code has now been run successfully on a range of accelerators including NVIDIA and A100 GPUs and for the first time on the newer AMD GPUs. We have also seen success in adapting the cloud scheme, and work will continue on other model physics components.
To achieve faster computer simulations of the Earth system with a lower energy footprint, we need our forecasting codes to be optimally mapped to heterogenous HPC architectures – something that we can only achieve through our collaboration with hardware architecture experts, software experts and domain specialists.
Efficient use of the emerging European high-performance computing platforms is fundamental to DestinE, in which ECMWF is working alongside EUMETSAT and ESA to create a digital replica of the Earth. Part of our work in the first phase is to develop the Digital Twin Engine, the complex software and hardware environment needed for very high-resolution modelling of the Earth system.
Within DestinE, we are beginning to run the IFS on some of the world’s top supercomputers across Europe, made available through the European High Performance Computing Joint Undertaking (EuroHPC), such as Meluxina in Luxembourg and the pre-exascale machines LUMI in Finland, Leonardo in Italy and MareNostrum5 in Spain. Such developments open up the possibility of distributed computing and data handling.
We have also achieved a major step toward kilometre-scale Earth system modelling with the world’s first kilometre-scale coupled atmosphere–ocean global simulations achieved with the CPU/GPU hybrid architecture of the Summit computer (under the US INCITE programme) and the JUWELS booster in Germany (under PRACE (Partnership for Advanced Computing in Europe)).
Our close collaboration with computer scientists, vendors, and the HPC and numerical weather prediction communities is crucial to progress.
ECMWF’s continuing leadership role enables us to support our Member and Co-operating States in their own digital goals, and to provide a forecasting system set to reap the benefits of the exascale age.
Background image: © gorodenkoff / iStock / Getty Images Plus