Energy Efficiency through Tuned Approximation

We present two algorithmic and software development campaigns of the Extreme Computing Research Center at KAUST targeting energy efficiency in dense (but data sparse) linear algebraic kernels. Savings are achieved while preserving application-grade accuracy by exploiting block low-rank representations and mixing precisions. These are implemented on mature accelerators, namely GPU and vector engines, and emerging accelerators for computational science and engineering, such as FPGAs and Wafer Scale Engines (WSEs). We include highlights of a 2022 Gordon Bell finalist in geospatial statistics and a 2023 Gordon Bell submission in seismic processing.