Optimized GPU implementation of Merck Molecular Force Field and Universal Force Field

Abstract

In silico research is important for many fields of modern science, from small molecule thermochemistry to drug design. Published algorithms and implementations all strived for exploiting the actual technology and hardware in the best and most efficient way. In this article we demonstrate novel GPU implementations of the well known Merck Molecular Force Field (MMFF94) and Universal Force Field (UFF) algorithms which are near to utilize the theoretical peak performance of the GPU the software runs on. A double-precision speedup of 55× for MMFF94 and 140× for UFF is achieved, with the factor being 90× for the single-precision implementation of MMFF94. Tests were carried out on the MMFF94 validation suite and on various length generated peptide chains.