HPC Limitations in Running Global Cloud-Resolving Weather Prediction Models

Recent estimates show that over half a million CPU processors would be needed to run the FV3GFS model at global cloud resolving (< 3KM) scales operationally at NCEP. Advances in processor technologies, improvement in model efficiencies including inter-processor communications, and I/O, are on-going and will likely decrease the number of processors to a degree, but the essential problem remains. It will be expensive to run models like the FV3GFS at cloud-resolving scales in operations. Newer generation model, under development such as the MPAS (by NCAR) and the LFRiC (by UKmet), may be more efficient and likely show further benefit using GPUs. Recent interest in cloud computing may offer cost improvement over traditional big-iron systems. This presentation will discuss these topics, show performance results, and assess challenges faced in running global, cloud resolving models in the future.