Energy-Efficient GPU Computing With Mixed-Precision Modeling for Climate/Weather Applications

Did you know that estimating spatial models is one of the main computational challenges in large geospatial data analysis? Traditionally, geospatial data is processed in FP64. The future of simulations is all about exploiting hardware features driven by the AI market for low-precision computations. Modern NVIDIA GPUs lead the charge with huge gains from low-precision computations that directly translate into energy savings. We harness their low-precision computing capabilities to introduce a mixed-precision geospatial modeling approach. Our adaptive precision conversion strategy, integrated into the ExaGeoStat software powered by the PaRSEC runtime, has turbocharged data modeling, achieving up to a 3x speedup against the full FP64 version while meeting the precision requirements of the application. Our approach is a game-changer for geospatial statisticians, bringing speed, energy efficiency, and precision to the forefront of computational statistics for environmental applications.