Researchers propose SeT-Diff, a diffusion-based foundational model for HPC telemetry and time-series

Researchers from the University of Bologna and other institutions propose SeT-Diff, a diffusion-based foundational model designed to address limitations in current machine learning approaches for high-performance computing (HPC) telemetry and time-series data. Unlike traditional models that rely on static subsets of anonymous, fixed-position sensor variables, SeT-Diff conditions the generative process on each sensor's semantic description. This decouples system dynamics from the structure of the dataset, enabling greater adaptability to changing tasks or sensor configurations.

The authors evaluate SeT-Diff on a real-world supercomputer dataset, reporting a Mean Absolute Error (MAE) of 0.0470 for reconstruction tasks. The model also demonstrates zero-shot permutation stability, maintaining accuracy with negligible degradation even when sensors are shuffled—a property not typically observed in conventional architectures.

Beyond reconstruction, a single pre-trained SeT-Diff model performs multiple tasks, including data imputation, forecasting, and virtual sensing. In thermal inference, the model achieves an MAE of 0.033, highlighting its versatility as a data-driven digital twin for HPC systems.

The work is presented as the first foundational model for compute node telemetry and time-series, positioning SeT-Diff as a potential paradigm shift for modeling complex interactions between workloads, environmental parameters, and physical metrics in data centers.