Berlin 2012 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 17: Poster Session
MM 17.76: Poster
Montag, 26. März 2012, 17:00–19:00, Poster B
Accelerating DFT and TDDFT Electronic Structure Calculations Using Graphics Processing Units — •Samuli Hakala1, Ville Havu1, Jussi Enkovaara2, Martti Puska1, and Risto Nieminen1 — 1Department of Applied Physics, School of Science, Aalto University, Espoo, Finland — 2CSC - IT Center for Science Ltd., Espoo, Finland
A Modern Graphics Processing Unit (GPU) is an efficient many-core stream processor suitable for throughput-orientated General Purpose (GP) parallel computations. Usage of GPGPU in scientific calculations has increased a lot in recent years. We have accelerated an electronic simulations software GPAW using GPUs. GPAW is a Density Functional Theory (DFT) program package based on the Projector Augmented Wave (PAW) method. Time-Dependent Density Functional Theory (TDDFT) is implemented in the linear response and time propagations schemes. Physical quantities are represented in a uniform 3D real space grid.
We have implemented GPU accelerated versions of the most numerically intensive parts in GPAW calculations using NVIDIA CUDA. Multiple GPUs and cluster nodes can be utilized with MPI using domain decomposition or by parallelizing over k-points. High performance is achieved by hand tuning the CUDA kernels and by minimizing data transfers between the GPU and the host computer. We describe our implementation and analyze the performance and the scalability of the code. Our results show that GPUs can provide significant speed-ups in both DFT and TDDFT calculations.