Graphics Processing Units (GPUs) have become the dominant type of processor on the TOP500 list of most powerful HPC systems. NOAA’s Global Systems Laboratory has been exploring GPUs since 2010. As part of the work, GSL developed the Non-hydrostatic Icosahedral Model (NIM) and demonstrated performance-portability of the model on CPUs, GPUs and MIC processors with a single source code. Performance results reported in 2015 showed GPUs to be about 3 times faster than CPUs with minimal changes to the code. The design and parallelization strategy was successfully adopted by the MPAS model, achieving similar performance characteristics. The successful work with the NIM led the GSL team to apply the same approach to the FV3 dynamical core used in the UFS today. The two year effort (2016 – 2018) to port the FV3 to GPUs was not ultimately successful however. This presentation will describe and compare the NIM and FV3 parallelization efforts, and share lessons learned in developing performance-portable codes.
