SC17 student cluster competition, Team Technical University of Munich and Friedrich–Alexander University Erlangen–Nürnberg: Reproducing vectorization of the Tersoff multi-body potential on the Intel Broadwell architecture

Abstract

Abstract Markus Hohnerbach et al., describe optimizations for calculating the potential of a bond between two atoms that depends on their neighboring atoms in “The Vectorization of the Tersoff Multi-Body Potential: An Exercise in Performance Portability”. In this paper we present the results of our team’s effort to reproduce the claims made by Markus Hohnerbach et al., on an Intel Broadwell architecture at the SC17 Student Cluster Competition. We first demonstrate that these optimizations result in no relevant loss of precision using types with different floating point representations and then compare their respective runtime. Lastly, we show the scalability of the double precision version of the algorithm by using an increasing number of threads.