Workshop: FTXS: Workshop for Faults, Trustworthiness, and eXplainability for AI Systems at Scale
Authors: Trupesh Patel and Purushotham V. Bangalore (University of Alabama)
Abstract: High-performance computing studies conclude that user error causes most job failures, while large-scale AI training reports attribute most interruptions to hardware. The two are rarely compared, because they publish failure evidence in incompatible forms: attributed causes, scheduler outcome labels, or bare event rates.
We survey 162 public datasets and reports, assign each an evidence verdict, and normalize the usable evidence into an eleven-class taxonomy tagging each failure type as a cause, a symptom, or an outcome label. The resulting denominator-disciplined maps cover 38 systems across 44 matrix columns, and include a GPU error-class characterization spanning five hardware generations.
Read on common footing the literatures agree: failure rates are similar but exposure is not, counting failures and weighting by compute give opposite priority orders, interconnect leads attributable GPU hardware failures, and the largest logged error class is application-induced. The survey table, taxonomy, and source quotations are released as a public artifact.
Back to FTXS: Workshop for Faults, Trustworthiness, and eXplainability for AI Systems at Scale Archive Listing
Back to Full Workshop Archive Listing