GPU acceleration of many-body perturbation theory methods in MOLGW with OpenACC

TL;DR

This paper demonstrates how GPU acceleration using OpenACC significantly speeds up MOLGW, a many-body perturbation theory code, enabling large-scale molecular excited-state calculations with improved efficiency.

Contribution

The authors implement GPU acceleration in MOLGW using OpenACC, achieving up to 9.7x speedup over CPU-based methods, facilitating large-system MBPT calculations.

Findings

Achieved up to 9.7x speedup with GPU acceleration.

Enabled large-system excited-state calculations.

Demonstrated effective use of OpenACC for scientific computing.

Abstract

Quasiparticle self-consistent many-body perturbation theory (MBPT) methods that update both eigenvalues and eigenvectors can calculate the excited-state properties of molecular systems without depending on the choice of starting points. However, those methods are computationally intensive even on modern multi-core central processing units (CPUs) and thus typically limited to small systems. Many-core accelerators such as graphics processing units (GPUs) may be able to boost the performance of those methods without losing accuracy, making starting-point-independent MBPT methods applicable to large systems. Here, we GPU accelerate MOLGW, a Gaussian-based MBPT code for molecules, with open accelerators (OpenACC) and achieve speedups of up to 9.7x over 32 open multi-processing (OpenMP) CPU threads.