Compile-Time Optimization of the Energy Consumption of Numerical Computations
Dorra Ben Khalifa, Matthieu Martel · 2024
Over the past decade, precision tuning has become one of the key techniques for achieving significant gains in performance and energy efficiency. This process consists of substituting smaller data types to the original data types assigned to floating-point variables in numerical programs in such a way that accuracy requirements remain fulfilled. In this article, we discuss the time and energy savings achieved using our precision tuning tool, POPiX. We validate our results on a set of numerical benchmarks covering various fields.