SC17 Denver, CO

P13: Large-Scale GW Calculations on Pre-Exascale HPC Systems


Authors: Mauro Del Ben (Lawrence Berkeley National Laboratory), Felipe H. da Jornada (University of California, Berkeley), Andrew Canning (Lawrence Berkeley National Laboratory), Nathan Wichmann (Cray Inc), Karthik Raman (Intel Corporation), Ruchira Sasanka (Intel Corporation), Chao Yang (Lawrence Berkeley National Laboratory), Steven G. Louie (Lawrence Berkeley National Laboratory; University of California, Berkeley)

Abstract: The accurate determination of excitation spectra of materials, such as the electronic band gap, is critical for the design of novel devices, including photovoltaics, transistors, batteries, and LEDs. Many-body perturbation-theory methods, and the ab-initio GW approach in particular, have emerged over the last decades as the gold standard for computing these quantities. However, the ab-initio GW formalism is often limited to systems of at most 100 atoms due to its computational complexity. We present here large scale GW calculations of crystalline defect problems, relevant for the performance of semiconductors, with up to 1000 atoms, on the Cori system at NERSC. We show that the GW method is particularly well suited for exascale/pre-exascale systems. Our implementation, which uses a combination of new algorithms and optimizations targeted at many-core CPU architectures, scales well to the entire Cori system, and obtains a significant fraction of peak performance.
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index