HONG KONG BAPTIST UNIVERSITY
FACULTY OF SCIENCE

Department of Computer Science Colloquium
2012 Series

GPU Implementation of a Plane Wave Pseudopotential Density Functional Theory Code: Petot

Dr. Lin-Wang Wang
Staff Scientist, Lawrence Berkeley National Laboratory
Berkeley
USA

Date: November 30, 2012 (Friday)
Time: 2:30 - 3:30 pm
Venue: RRS638, Sir Run Run Shaw Building, Ho Sin Hang Campus

Abstract
One rapidly developing area in scientific computing is the use of general purpose graphic processing units (GPU). In materials science simulations, the plane wave pseudopotential method is the most widely used density functional theory (DFT) simulation method. We have implemented our plane wave pseudopotential code PEtot using GPU. This requires a redesign of the parallelization algorithm. In particular, we calculate the band to band overlapping matrix elements in G-space parallelization, with the wave function to Hamiltonian application in band index parallelization. We found that it is possible to accelerate the original MPI code by a factor of 20 by the GPU code. As a result, in a molecular dynamics (MD) simulation of a 512 atom GaP system, each MD step takes only about 10 seconds using 256 GPUs.

Biography
Dr. Wang has 20 years of experience in large scale electronic structure calculations. He has worked on O(N) electronic structure calculations in the early 1990s. Working with Alex Zunger, he invented the folded spectrum method which pushed the limit of non-self-consistent electronic structure calculations from 100 atoms to thousands of atoms. He developed a linear combination of bulk bands (LCBB) method for semiconductor heterostructure electronic structure calculations, which allows the calculation of million-atom devices. He developed generalized moments method which calculates the density of states and optical absorption spectra of a given system without explicit calculation of its eigenstates. He also developed a popular parallel total energy plane wave pseudopotential program (PEtot). He invented a charge patching method, which enables ab initio accuracy in thousand atom calculations for nanosystems. He has developed a linear scaling three dimensional fragment method (LS3DF), which can be used to self-consistently calculate systems with tens of thousands of atoms. Recently, he developed a non-adiabatic molecular dynamics method which can simulate a 5,000 atom system for 10 picoseconds.

*This colloquium is jointly organized by Department of Computer Science / Mathematics / Physics / Institute of Computational and Theoretical Studies of Hong Kong Baptist University.

********* ALL INTERESTED ARE WELCOME ***********
(For enquiry, please contact Computer Science Department at 3411 2385)

http://www.comp.hkbu.edu.hk/v1/?page=seminars&id=217&lang=sc
Photos