How to use VASP 6.2.0 on swing.

We have built the open_acc version of VASP.6.2 for swing. You will generatlly find your calculations will run several times faster on GPUs than on a broadwell node. The speedup you get will depend on the size of your model and the values of KPAR and NSIM in your INCAR file. See for details. We have used the silicaIFPEN benchmark to get typical speedups. The silicaIFPEN benchmark uses the PBE functional for a supercell model of a defect in cristobalite containing 240 atoms. A single GPU was 2.94 times faster than a broadwell node, two gpus are 60% faster than one GPU, and four gpus are 60% faster than two GPUs. Eight GPUs are 10% faster than four GPUS. We have not seen good scaling for more than four GPUs for this benchmark.

Not all of VASP has been ported to GPUS. Semi-Local and hybrid functionals has been ported. The hybrid funtionals should be very fast on the GPUs. RPA is still a work in progress and it does not run significantlly faster on the GPUs than on the broadwells.

An example script can be found in /soft/vasp/6.2.0/swing/