GPGPU for orbital function evaluation with a new updating scheme

Abstract

We have accelerated an ab initio quantum Monte Carlo electronic structure calculation using general purpose computing on graphical processing units (GPGPU). The part of the code causing the bottleneck for extended systems is replaced by Compute Unified Device Architecture-GPGPU subroutine kernels which build up spline basis set expansions of electronic orbital functions at each Monte Carlo step. We have achieved a speedup of a factor of 30 for the bottleneck for a simulation of solid TiO(2) with 1536 electrons. To improve the performance with GPGPU we propose a new updating scheme for Monte Carlo sampling, quasi-simultaneous updating, which is intermediate between configuration-by-configuration updating and the widely used particle-by-particle updating. The error in the energy due to by the single precision treatment and the new updating scheme is found to be within the required accuracy of ~10(-3) hartree per primitive cell.