SeT-Diff Creates Semantic Foundation Model for HPC Telemetry

A new research paper unveils SeT-Diff, a pioneering foundation model designed for High-Performance Computing (HPC) telemetry and time-series data. Traditional machine learning methods for data centers often rely on static sensor variables, becoming obsolete when tasks or sensor configurations change. SeT-Diff addresses this limitation by decoupling system dynamics from dataset structure. This diffusion-based model conditions its generative process on the semantic descriptions of each sensor. This innovative approach allows SeT-Diff to maintain accuracy even when sensor metrics vary or are shuffled, demonstrating zero-shot permutation stability. It effectively creates flexible digital twins capable of modeling the intricate relationships between workloads, environmental factors, and physical metrics. Experiments on real-world supercomputer data show SeT-Diff achieving a Mean Absolute Error (MAE) of 0.0470 in reconstruction tasks. A single pre-trained model can perform multiple functions, including data imputation, forecasting, and virtual sensing, achieving a 0.033 MAE in thermal inference. This makes SeT-Diff a highly effective and adaptable data-driven digital twin for complex HPC systems.