MPAS and GPU-accelerated computing architectures

In collaboration with the University of Wyoming and NVIDIA, The Weather Company, an IBM subsidiary, has funded NCAR to adapt the Model for Prediction Across Scales (MPAS) to run on GPU-accelerated computing architectures. Through this work, MPAS-Atmosphere is able to run roughly seven times faster on compute nodes with two POWER9 CPUs and four NVIDIA Volta 100 GPUs compared to nodes like those in NCAR’s current flagship computer, Cheyenne. The accelerated model is now running operationally at The Weather Company, and it represents one of the first uses in the world of a global, GPU-accelerated model for publicly available weather forecasts.

Looking toward the future, this work introduces generality to the MPAS model source code that will prepare MPAS – both global and regional – for the next generation of super-computers, which will take advantage of accelerators and heterogeneous architectures. The GPU-enabled model source code will be made freely available, enabling industry, academic researchers, and interested individuals to benefit from this work.

The simulation rate achieved by the accelerated MPAS-Atmosphere model on a computer with two POWER9 CPUs and 4 NVIDIA Volta 100 GPUs per node
Figure: The simulation rate achieved by the accelerated MPAS-Atmosphere model on a computer with two POWER9 CPUs and 4 NVIDIA Volta 100 GPUs per node is roughly seven times higher than the simulation rate achievable on CPU-only nodes like those in NCAR’s current flagship computer, Cheyenne. The simulation in this benchmark made use of a variable-resolution mesh with a global grid spacing of 15 km with parts of the mesh refining to a grid spacing of 3 km. Figure courtesy of IBM / The Weather Company.