With new supercomputer clusters supporting multiple GPU types, often from multiple vendors, maintaining software capable of using these resources becomes difficult. This talk explores the current state of the art with respect to portable yet still performant code, striving for the ideal of a single codebase that achieves good performance on all platforms.
Slides will be available for download here after the presentation.
Josh H. Davis is a third-year PhD student and NSF Graduate Fellow advised by Prof. Abhinav Bhatele in the Parallel Software and Systems Group at the University of Maryland. He completed his undergraduate studies in Computer Science and Philosophy at the University of Delaware in 2021. His main research interests center on improving developer productivity in high-performance computing, including performance portability, programming models, compilers, machine learning applications, and model checking. His current projects include evaluating performance portability in GPU programming models, automated formal verification of scalable data structures using model checking, and testing large language models for HPC programming tasks.