GPU implementations of some many-body potentials for molecular dynamics simulations

We study the efficiency of OpenCL implementations for Tersoff and embedded-atom interatomic potentials. We show that Tersoff potential can be computed faster using atomic operations rather than using longer kernel code. On the contrary better performance can be gained for embedded-atom potential without atomic operations. Numerical force computation algorithm is the slowest but shows the best scaling with the highest GPU acceleration. The GPU acceleration of different algorithms was evaluated and compared to the serial implementations of similar algorithms. The performance of GPU implementations is superior to their serial counterparts and depends on the algorithm and arithmetic precision. The corresponding benchmarks and performance comparison were done using NVidia GPUs and Intel CPUs.