Many-body perturbation theory (MBPT) is a powerful method to simulate electronic excitations in molecules and materials. However, its application to large systems has been limited by the high computational cost. To address this challenge, we developed the WEST code (http://west-code.org), a massively-parallel, GPU-accelerated implementation of MBPT that overcomes common computational bottlenecks such as summing over empty electronic states and inverting large dielectric matrices. WEST achieves excellent performance and scalability on various supercomputers by using a hierarchical parallelization strategy that minimizes CPU-GPU, GPU-GPU, and CPU-CPU data communications, overlap between computations and communications, and mixed precision in selected portions of the code. The GPU version of WEST achieves a substantial speedup compared to its CPU version, and scales efficiently to over 20,000 GPUs. Finally, we present large-scale full-frequency G0W0 calculations for realistic nanostructures and interfaces comprising up to 10,368 electrons, demonstrating the potential of GPU acceleration in enabling the efficient simulation of large systems with MBPT.
Alvaro Vazquez Mayagoitia