AI technologies, especially large language models, are transforming software development, but high-performance computing (HPC) software poses distinctive requirements around performance, portability, parallelism, and trustworthiness. This paper outlines the challenges of applying state-of-the-art AI to this specialized software domain.
We identify research directions for making HPC software development more productive and trustworthy with AI assistance, drawing on the goals of two U.S. Department of Energy projects for advancing HPC software via AI: Ellora and Durban.
@inproceedings{teranishi:hpc-ai-directions,
title = {Leveraging {AI} for Productive and Trustworthy {HPC} Software: Challenges and Research Directions},
author = {Keita Teranishi and Harshitha Menon and William~F. Godoy and Prasanna Balaprakash and David Bau and Tal Ben-Nun and Abhinav Bhatele and Franz Franchetti and Michael Franusich and Todd Gamblin and Giorgis Georgakoudis and Tom Goldstein and Arjun Guha and Steven~E. Hahn and Costin Iancu and Zheming Jin and Terry Jones and Tze Meng Low and Het Mankad and Narasinga Rao Miniskar and Mohammad Alaul Haque Monil and Daniel Nichols and Konstantinos Parasyris and Swaroop Pophale and Pedro Valero-Lara and Jeffrey~S. Vetter and Samuel Williams and Aaron Young},
year = {2025},
booktitle = {International Conference on High Performance Computing ({ISC} High Performance Workshops)},
}