A highly efficient implementation of direct GPU to GPU communication in a library for direct exchange calculations
PI: Stefan Goedecker (University of Basel)
April 1, 2015 - March 31, 2017
The strive for higher accuracy in electronic structure calculations is the dominating theme in density functional theory. Even though higher accuracy is in principle desired in any application it is particularly urgently needed in structure prediction and reaction pathway searches. Hybrid functionals that contain a certain fraction of direct exchange are at present considered to be among the most accurate functionals. The widespread use of hybrid functional calculations is at present hampered by their high numerical cost if they are used in combination with a systematic basis set such as plane waves or wavelets. This project proposes therefore a highly efficient implementation of direct exchange within density functional calculations using systematic basis sets by the use of GPUs and direct MPI communication between these GPUs. This will allow to recover the exact energies of the corresponding functional without any basis set limitations while obtaining at the same time the highest possible speed. The library will be open software and will in particular have interfaces to the BigDFT and Quantum-Expresso codes. It is to be expected that by these developments hybrid functional calculations, that were up to now mainly done in calculations using Gaussian basis sets, will also become much more widespread in calculations using systematic basis sets.