Quantum Chemistry on GPGPU

Marcin Makowski,1,2 Grzegorz Mazur,1,2 Piotr Kuzniarowicz,1,3

Roman Lazarski,1,4 Alina Mrenca,5 Krzysztof Kolasinski5

1Faculty of Chemistry, Jagiellonian University

2Academic Computer Centre CYFRONET AGH

3Department of Molecular and Material Sciences, Kyushu University

4Institute of Material Science and Material Technology, F. Schiller University

5Faculty of Physics and Computer Science, AGH University of Science andTechnology

KU KDM 20131 z 16

Plan

Introduction

GPGPU

niedoida

2 z 16

Plan

Introduction

GPGPU

niedoida

3 z 16

POWIEW

HPC Infrastructure for Grand Challenges of Science andEngineering

I ICM, CYFRONET, PCSS

I Acquisition of modern HPC architectures allowing forlarge-scale calculations

I Research groups working optimal computational models forHPC

3 z 16

POWIEW in CYFRONET

I HardwareI GPGPUI vSMPI cluster of fat nodes

I Software

I Software development

I Workshops

4 z 16

Computational Chemistry

ZEUS

I 34% of gridworkload

I 56% of localworkload

5 z 16

Plan

Introduction

GPGPU

niedoida

6 z 16

GPGPU

6 z 16

GPGPU in TOP100

VI 2008 XI 2008 VI 2009 XI 2009 VI 2010 XI 2010 VI 2011 XI 2011 VI 20120

5

10

15

20

25

7 z 16

GPGPU in CYFRONET

I HardwareI 48 NVIDIA Tesla M2050 GPUs (24 servers)I 160 NVIDIA Tesla M2090 GPUs (20 servers)I almost 140 TFlops

I SoftwareI TeraChem (quantum chemistry)I NAMD (molecular mechanics)

I Software development

8 z 16

Plan

Introduction

GPGPU

niedoida

9 z 16

niedoida

I General-purpose computational quantum-chemistry package

I Created and developed at Faculty of Chemistry (JagiellonianUniversity) in collaboration with CYFRONET

I Optimal implementation of existing methods

I Development of original, efficient methodologiesI Aiming at large systems

I advanced hybrid parallelizationI GPGPU computing

9 z 16

Testbed

I Hardware

I CPU: Intel Xeon X5670

I GPUs: Tesla M2050, GeForce 550 TiI Software

I niedoidaI TeraChem (ver. 1.50K)

I Molecular systemsI polyglycine chainsI 6-31G basis setI RHF and B3LYP calculations

10 z 16

GPU vs CPU: Hartree-Fock

2

4

6

8

10

12

5 10 15 20 25 30 35

Sp

eed

up

System size (glycine units)

GeForce GTX 550 TiTesla M2050

11 z 16

GPU vs CPU: B3LYP

2

4

6

8

10

12

5 10 15 20 25 30 35

Sp

eed

up

System size (glycine units)

GeForce GTX 550 TiTesla M2050

12 z 16

Multi-GPU efficiency

1

1.2

1.4

1.6

1.8

2

2.2

2.4

5 10 15 20 25 30 35

Sp

eed

up

System size (glycine units)

JKGrid

13 z 16

MethodSpeedup

vs CPU vs TeraChem

HF 11.10 0.30B3LYP 8.94 0.37

grid integration 2.33 0.69

14 z 16

Future

Projects in progress/planned:

I integration for d- and f-type basis functions

I RI-DFT and RI-MP2

I CIS and TD-DFT excited states calculationsI gradients and second derivatives:

I geometry optimizationI Born-Oppenheimer molecular dynamicsI vibrations

I hybrid parallelization

15 z 16

To Be Continued. . .

This work was carried out thanks to the”POWIEW”POIG.02.03.00-00-018/08-00 project.

This project is co-funded by the European Regional Development Fund(ERDF) as a part of the Innovative Economy program.

16 z 16