Excited-state calculations on exascale supercomputers using BerkeleyGW

The GW method is widely used to describe excited-state phenomena in materials. The GW method displays higher computational cost than less accurate approaches but its computational kernels are better suited to massive parallelization and GPU acceleration. In this talk, we present GPU optimizations that are implemented in the open-source BerkeleyGW software package. Additionally, we demonstrate portability among different system architectures (NVIDIA, AMD, Intel) while maintaining a high fraction of peak performance. The code can scale up to thousands of GPU nodes on leadership class HPC architectures with excellent strong and weak scaling. We will discuss various performance portability strategies of BerkeleyGW's most computationally intensive kernels when using different GPU programming models across multiple GPU hardware.