Optimization of a Numerical Weather Prediction Dynamical Core on the INTEL Many Integrated Core Computing Architecture: An Update on Options, Methods, Results and Future Efforts

This paper investigates the optimization of Numerical Weather Prediction (NWP) model dynamical core computations on the INTEL Many Integrated Core Architecture. The goal of this work is to identify, test and verify optimization methods that can be used in operational weather prediction models. The work focuses on single MIC optimization and leaves the multiple MIC optimizations for later study. The typical finite difference grid point weather prediction model runs at a low percentage of peak performance on scalar architecture and this work focuses on maximizing single card performance. Optimization methods will be applied to test computations and the results and extension to operational models will be discussed.