PRISM: Dynamic Primitive-Based Forecasting for Large-Scale GPU Cluster Workloads

TL;DR

PRISM is a novel forecasting framework that uses primitive-based decomposition and spectral refinement to accurately predict volatile, heterogeneous GPU workloads, improving resource management in AI infrastructure.

Contribution

It introduces a dual representation approach combining dictionary-driven decomposition with spectral refinement for workload forecasting.

Findings

Achieves state-of-the-art forecasting accuracy on large-scale GPU traces

Reduces burst-phase errors significantly

Provides a robust foundation for dynamic GPU resource management

Abstract

Accurately forecasting GPU workloads is essential for AI infrastructure, enabling efficient scheduling, resource allocation, and power management. Modern workloads are highly volatile, multiple periodicity, and heterogeneous, making them challenging for traditional predictors. We propose PRISM, a primitive-based compositional forecasting framework combining dictionary-driven temporal decomposition with adaptive spectral refinement. This dual representation extracts stable, interpretable workload signatures across diverse GPU jobs. Evaluated on large-scale production traces, PRISM achieves state-of-the-art results. It significantly reduces burst-phase errors, providing a robust, architecture-aware foundation for dynamic resource management in GPU-powered AI platforms.