One of the major strengths of the GFDL Finite-Volume Cubed-Sphere Dynamical Core (FV3) is its efficiency and scalability on traditional CPUs. FV3 was designed to fit the MPI-OpenMP paradigm which has served HPC well for three decades. But this may be changing as emerging exascale systems are using manycore processors, especially GPUs. Even newer architectures like ARMs and FPGAs are on the horizon. Previous results to re-write FV3 for GPUs have shown 10x or more speedups in the entire core. More recent results have shown 50x GPU speedups to some modules within FV3. However, each new architecture requires re-writing of scientific codes for best performance. We are collaborating with Vulcan Inc, NASA Goddard, and university partners to port FV3 and the UFS into the GT4py Domain-Specific Language. By doing so the model can be re-compiled for target processors in a way so the code and data can be laid out for optimum performance. Progress and prospects for both traditional and DSL ports will be discussed.
