high performance computing on graphics processing units: hgpu.org

Advancements in massively parallel sampling of the conformational space of biomolecules enables, for example, protein structure prediction, in-silico drug development and cell signaling. Despite the existence of highly distributed protein simulation architectures like POEM@HOME, there was no abundant computational resource both strong and serial strength and in parallel sampling. In this study we investigate the optimization of our N-body Lennard-Jones force field for the efficient Monte-Carlo sampling of small to medium-size biomolecules on massively parallel architectures, like modern GPUs. We benchmark both NVIDIA and AMD GPU chipsets in the OpenCL framework on comparison to CPU architectures. The N-body interactions are broken down into small local grids, which fit into the local GPU caches to permit simultaneous evaluation. Using the N-body term we accelerate the Lennard-Jones and Clash-Potential of the complete free-energy PFF02 [1] shown to fold a multitude of different protein-folds and implement a modified structure-based Lennard Jones force field. We proof the applicability of our novel force field by reversible folding-simulations of a three-helix protein using this Go-potential from completely unifolded structures.