Workshops Archive

Mechanism-Aware Analysis of Cross-Platform GPU Fault-Injection Outcomes


Workshop: Fourth International Workshop on HPC Testing and Evaluation of Systems, Tools, and Software (HPCTESTS 2026)

Authors: Maisy Dunlavy, Zhiling Lan, and Michael Papka (University of Illinois at Chicago)

Abstract: HPC test practitioners frequently compare fault injection results across GPU platforms, software stack versions, and computing facilities. Standard outcome categories such as masked faults, silent data corruptions (SDCs), and detected unrecoverable errors (DUEs) provide useful campaign summaries, but matching labels or distributions do not necessarily indicate comparable behavior. We present a mechanism aware method for interpreting cross platform GPU fault injection results and apply it to six reproducible cases on production NVIDIA, Intel, and AMD GPU systems. The cases show that bit position can determine whether a test completes with SDC or causes device failure, that aligned injections can become different effective faults after backend lowering, and that faults reaching control or address logic may remain masked. We recommend retaining lightweight runtime evidence for all tests and performing targeted propagation analysis on key cases before declaring cross-platform results equivalent.


Back to Fourth International Workshop on HPC Testing and Evaluation of Systems, Tools, and Software (HPCTESTS 2026) Archive Listing

Back to Full Workshop Archive Listing