Performance Profiling on AMD GPUs – Part 1: Foundations
Auto-discovered
Visit resource ↗Description
This is Part 1 of a three-part ROCm Blogs series by Gina Sitaraman, Thomas Gibson, Luka Stanisic, Giacomo Capodaglio, Alessandro Fanfarillo, and Asitav Mishra on performance profiling foundations for AMD GPUs.
It defines novice and advanced audiences, lists prerequisites such as single-GPU awareness versus multi-GPU and roofline fluency, and surveys ROCm profiling building blocks including rocprofiler-sdk, rocprofv3 for traces and counters, rocprof-sys for host and device traces, and rocprof-compute for kernel analysis and rooflines. It also outlines installation and testing via AMD HPCTrainingDock and HPCTrainingExamples scripts and modules.
It is for HPC and GPU developers who will profile on AMD Instinct hardware, from first-time profilers to engineers tuning multi-GPU runs. Basic knowledge that kernels run on a GPU is enough for the novice track.
Access is a free HTML blog post on AMD's ROCm blogs site, dated June 26, 2025, as the series introduction before later novice and advanced installments.
It is useful as a map of current AMD profiling tools and setup before deeper optimization posts.