Revisting Seminumerical Methods for Electronic Structure Calculations in the Age of Exascale Computing

With the increasing reliance on the use of GPGPU accelerators in modern

high-performance computing, the exploitation of these new and emerging

architectures has become a persistent challenge for electronic structure

methods developers. This challenge has recently been exacerbated by an

increasing diversity in the accelerator architectures and associated

programming models available on contemporary (pre-)exascale computing

resources, including GPGPU hardware from NVIDIA, AMD and Intel. Due to the

drastically different concurrency and memory models used in CPU and GPGPU

architectures, it has often been the case that the traditional strategies used

for compute intensive operations on the CPU are not necessarily the optimal

choice for GPGPU architectures. In this talk, we examine recent work in the

development of numerical and semi-numerical techniques for the construction of

the Fock matrix in hybrid Kohn-Sham density functional theory which have been

demonstrated to exhibit excellent performance on contemporary GPGPU

architectures and prove promising for application on massively parallel

(post-)exascale computing resources. Further, we present a modular design

pattern for these methods which allows for the simultaneous targeting of

several accelerator architectures in a single software infrastructure.