Facilitating Electronic Structure Calculations on GPU-based Exascale Platforms

The IDEAS Productivity project, in partnership with the DOE Computing Facilities of the ALCF, OLCF, and NERSC, and the DOE Exascale Computing Project (ECP), organizes the webinar series on Best Practices for HPC Software Developers.

As part of this series, we offer one-hour webinars on topics in scientific software development and high-performance computing, approximately once a month. The April webinar is titled Facilitating Electronic Structure Calculations on GPU-based Exascale Platforms; and will be presented by Jean-Luc Fattebert (Oak Ridge National Laboratory). The webinar will take place on Wednesday, April 12, 2023, at 1:00 pm ET.

Abstract:

GPUs accelerators offer the prospect of speeding up ab initio molecular dynamics and other large-scale first-principles atomistic simulations. Taking advantage of these devices is, however, not a trivial task given their specificities. Some algorithms struggle, while others thrive with the high level of thread concurrency available on modern GPUs. The PROGRESS and BML libraries, developed within ECP’s Co-design Center for Particle Applications (CoPA) project, allow electronic structure codes to offload their most expensive kernels, with a unified interface for various matrix formats and computer architectures. The webinar will focus on implementations and algorithmic choices made in those libraries, and lessons learned while trying to achieve performance portability on exascale platforms. Specifically, the webinar will discuss eigensolvers and their alternatives, as well as strong scaling in fast time-to-solution in molecular dynamics.