Workshops Archive

Adaptive Block Mixed-Precision Cholesky Factorization for Large-Scale Space-Time Gaussian Processes


Workshop: High Performance Computing for Environmental and Earth Sciences (HPC4EES)

Authors: Mary Lai Salvana (University of Connecticut)

Abstract: Gaussian process models underpin many environmental applications but are limited by the dense Cholesky factorization of an $N \times N$ covariance matrix that costs $O(N^3)$ operations and $O(N^2)$ memory. Low-precision arithmetic can reduce these costs, yet naive \fp{8}, half-, or single-precision Cholesky may fail on ill-conditioned covariance matrices because rounding destroys positive definiteness. We develop an adaptive block mixed-precision Cholesky factorization that partitions the covariance into blocks and assigns each the lowest precision satisfying a prescribed accuracy budget. A condition-number safety factor guarantees completion by preserving positive definiteness after rounding. Applied to a space-time Gaussian process for forecasting storm wind gusts over the northeastern United States, an extreme-weather forecasting task critical to hazard warning, the method scales to 131,589 observations on a single GPU, performs most arithmetic in \fp{8}, preserves the log-likelihood, and makes maximum-likelihood estimation practical at a scale beyond double precision while producing calibrated forecasts with predictive uncertainty.


Back to High Performance Computing for Environmental and Earth Sciences (HPC4EES) Archive Listing

Back to Full Workshop Archive Listing