Large-Scale GW Calculations on Pre-Exascale HPC Systems*

Large-scale GW calculations are required to accurately describe excited state phenomena in materials, which is critical for the design of novel devices in many fields. However, application of the GW method to complex systems is often limited due to the high computational cost. Reduced time to solution can be achieved through novel methods, algorithms and optimal implementations on modern HPC systems. We demonstrate these capabilities utilizing a highly-optimized version of the BerkeleyGW software package for HPC many-core architectures. The developed code, tested on Cori@NERSC (a Cray XC40, Xeon-Phi powered system), is capable of scaling to the full-machine, using a high fraction of peak performance and achieving excellent time to solution for systems of thousands of atoms. A high fraction of peak and good parallel scaling comes from an improved data layout where the computationally intensive work at the node level is cast as large ZGEMM operations, combined with a ring-based communication scheme, which avoids collective operations and allows overlapping with computation.