Hacia el uso de procesadores gráficos (GPUs) en simulaciones Híbridas Cuántico-Clásicas (QM-MM)

GONZALEZ LEBRERO MC; NITSCHE, MATÍAS

La Falda, Córdoba

Congreso; PRIMER ENCUENTRO NACIONAL DE COMPUTACION DE ALTO RENDIMIENTO PARA APLICACIONES CIENTiFICAS; 2010

Universidad Nacional de Córdoba: Facultad de Cs. Químicas y Facultad de Matemática, Astronomía y Física y el CONICET: Instituto de Astronomía y Física del Espacio, Instituto de Investigaciones Físicoquímicas de Córdoba, Instituto de Física Enrique Gaviola

The mayor cost in a QM-MM simulation resides in the calculation of the electronic structure of the quantum region. For this, the optimization of this part for the calculation shifts the limits in the accuracy and size of the system who can be studied. This work presents an implementation of electronic structure calculations based on the Density Functional Theory (DFT), specifically designed to be executed on a graphics processor (GPU). The software MOLECOLE was used as starting point of this development. Following the literature on this subject, the known efficient algorithms are implemented. However, novel adaptations that take into account both the characteristics of the used hardware, and the particular types of molecular systems in which this work focuses (QM-MM), are introduced. The appeal of such hardware is its great computational power, which is a product of the highly parallelized architecture implemented on it. In calculations where this parallelism can be exploited, it is possible to obtain signifcant improvements in execution times. With the software developed in this work an implementation which is at par with the known Gaussian ´03 in terms of numerical quality, but with a reduction in computing time to about ten times (10x) is achieved, by comparing to a serial execution, and to four times (4x) by comparing to a parallelized execution over four CPU cores. Other comparisons are also made between the original software Molecole, SIESTA and an implementation that targets conventional processors (CPU), but analogous to the one developed for GPU (see Figure 1). Finally, it is to be expected that the experience gained through solving the various problems encountered when working on this type of hardware could be generalized to future projects. Moreover, the results not only motivate to advance with the present work, but also demonstrate the usefulness of these processors to speed up these type of calculations and possibly other similar.