Workshops Archive

Deriving Vendor-Agnostic GPU Metrics from AMD MI300A and NVIDIA H100 Hardware Counters


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

Authors: Amr Abouelmagd (Tennessee Tech University), Michael McKinsey and David Boehme (Lawrence Livermore National Laboratory), Stephanie Brink (Lawrence Livermore National Laboratory (LLNL)), and Olga Pearce (Lawrence Livermore National Laboratory)

Abstract: Large-scale scientific applications increasingly run on GPUs from multiple vendors, and comparing performance across architectures is essential to porting and tuning codes. Such comparisons rely on hardware counters, but AMD and NVIDIA expose counter sets designed independently, with different names, granularities, and semantics, and no correspondence between them exists. We construct and validate a mapping between AMD MI300A and NVIDIA H100 hardware counters for FP64 compute and memory-hierarchy metrics, scoring each candidate pair across 12 kernels from the RAJA Performance Suite. Counters that appear comparable can measure different quantities: AMD reports FP64 instructions per wavefront without indicating how many of its 64 threads were active, which makes it report up to 3.19× more FP64 FLOPs than NVIDIA for the same kernel. We show how to correct this, and report every mapping we tested, including those that agree poorly and the metrics for which no mapping exists.


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

Back to Full Workshop Archive Listing