PRISM: Probabilistic Runtime Insights and Scalable Performance Modeling for Large-Scale Distributed Training

Large model training beyond tens of thousands of GPUs is an uncharted territory. At such scales, disruptions to the training process are not a matter of if, but a matter of when -- a stochastic process degrading training productivity. Dynamic runtime variation will become increasingly more frequent as training scales and GPUs are operated in increasingly power-limited and thermally-stressed environments. At the 64,000+ GPU scale, we already observe 12% variability for frontier foundation model training. Motivated by our analysis and the large design space around performance variability, we present PRISM -- a performance modeling framework that captures the stochastic nature of large-scale distributed training. The core of PRISM is a statistical model that composes operator-level latency distributions through workload dependencies. Across 14 diverse training configurations spanning hundreds to 64K+ GPUs, PRISM estimates p95 execution time within 5.4% error. Using PRISM, we explore the design and optimization space of distributed training, enabling principled, variability-aware recommendations that can improve performance and system efficiency at scale.