Abstract

The use of density functional theory (DFT) simulations has proved to be a powerful tool in the Earth and Environmental Sciences. However, these calculations require a large computational power and are limited to ~1000 atoms in big supercomputers. An alternative to parallelize DFT codes in CPU supercomputers is the use of GPU cards. These contain a large number of threads that can accelerate codes which are properly programmed for parallel calculations. In this mini-review we have evaluated which factors are crucial to obtain an appropriate acceleration in the process of moving CPU codes to their GPU version: memory transfer, work flows and CPU/GPU ratio. Accelerations up to 20-40 times the pure CPU version of the DFT code have been achieved. This makes that the additional cost of GPUs cards is less than the price/performance obtained.