Application of Machine Learning Techniques to Characterize AI Benchmarks Using Hardware Events

Nowadays, in a dynamic data center landscape, effective distribution of computational resources is crucial to achieving peak performance and cost-efficiency. This study addresses the challenge of distinguishing artificial intelligence work-loads from other tasks within data centers, focusing on optimizing their execution on specialized AI accelerators like GPUs. By integrating machine learning algorithms with hardware performance counters, the study profiles AI workloads and develops predictive models. These models enable accurate detection and optimization of AI compute tasks, aiming to enhance resource allocation and overall performance in modern data center environments. The study achieves 100% precision in distinguishing AI / ML compute workloads from other benchmarks in the proposed experiments, underscoring its effectiveness in improving data center efficiency and productivity.