Porting the LAMMPS/EAM benchmark to OpenCL TM with limited code modifications

Introduction. Demonstrating the use of GPUs for technical computing normally focuses on the use of GPU-tuned algorithms for accelerating applications over the baseline performance on a CPU. Although speedups of as much as 100x or greater are sometimes reported, more realistic speedups are in the range of 2x to 10x when careful consideration and effort is given to comparable optimizations for the CPU. The acceleration of GPU-optimized code often neglects the realities of existing production HPC codes in terms of acceptable modifications, where absolute performance is only one of many factors considered in software design decisions. This white paper discusses the initial results from an investigation of a pre-defined benchmark for a production molecular dynamics code (LAMMPS) using OpenCL