Building a Productive AI-Assisted High Performance Computing Software Ecosystem
Abstract: Developing and maintaining scientific software is becoming increasingly challenging as high performance computing (HPC) systems grow more heterogeneous. Large language models (LLMs) offer a promising way to improve developer productivity, but current models still struggle with HPC code because of limited training data, large codebases and the need to reason about program structure and behavior beyond source text alone.
In this talk, we will describe our work on improving LLMs for HPC software development, including long-context reasoning, richer code representations, tool-using agents and evolutionary search for code optimization and transformation. We will discuss how these approaches can help with software-development tasks and reduce the expert effort needed to adapt scientific codes to evolving HPC systems.
Bio: Harshitha Menon is a research scientist at the Center for Applied Scientific Computing at the Department of Energy’s (DOE) Lawrence Livermore National Laboratory. She is the principal investigator of Ellora, a DOE project advancing LLM approaches for HPC software development. Harshitha received her Ph.D. in 2016 and M.S. in 2012, both from the University of Illinois at Urbana-Champaign. She is a recipient of the Association for Computing Machinery/Institute of Electrical and Electronics Engineers-Computer Science George Michael Fellowship in 2014, the Anita Borg Scholarship in 2014 and was selected for Lawrence Livermore’s Director’s Early and Mid-Career Recognition Award in 2026.
Series: See all upcoming talks at https://www.anl.gov/mcs/lans-seminars.