Workshops Archive

Accelerating HPC Performance Analysis with AI Agents


Workshop: 8th Workshop on Programming and Performance Visualization Tools (ProTools)

Authors: Kevin Huck (Advanced Micro Devices (AMD))

Abstract: Successful performance analysis typically involves a combination of established workflows, measurement tools, architectural knowledge, performance models, hardware counter analysis, parametric studies, and visualization. AI agents can integrate and automate many of these activities, significantly reducing the effort required to collect, organize, analyze, and interpret performance data. Agents are particularly effective at automating command-line data collection, managing results from parametric studies, applying performance models to known architectures, interpreting well understood hardware counter data, and performing static analysis of compiler reports and generated assembly code. In this talk, we will present several success stories that demonstrate how AI agents can be used to analyze HPC proxy applications and accelerate the performance engineering process. For example, analysis of the RAJAPerf suite on AMD MI300A GPUs identified optimization opportunities that produced a 34× speedup in one kernel. We will also discuss practical lessons learned, including common challenges and limitations encountered when working with AI agents. We will also discuss emerging opportunities for agent-assisted performance analysis, including the skills, interfaces, and rules that performance measurement and analysis tool developers should consider providing to maximize agent effectiveness. Finally, we will highlight AI agent capabilities being developed for current and future releases of the ROCm ecosystem.


Back to 8th Workshop on Programming and Performance Visualization Tools (ProTools) Archive Listing

Back to Full Workshop Archive Listing