Improving Energy Efficiency in Scientific and AI Computing

Abstract: Energy efficiency has become a central challenge in leadership computing as applications increasingly rely on powerful and heterogeneous accelerator platforms. This seminar presents an overview of Ph.D. research developed through collaboration with the Argonne Leadership Computing Facility. The work investigates how application behavior, resource utilization and performance requirements can be leveraged to improve energy efficiency without compromising time-to-solution.

Through studies conducted on Polaris and Aurora, the research examines how accelerator resources can be used more effectively across scientific computing and artificial intelligence applications. The topics include graphics processing unit (GPU) resource sharing, application co-execution and performance prediction. Although these techniques can improve resource utilization and reduce energy consumption, their effectiveness depends strongly on workload characteristics, system architecture and execution objectives.

The results show that energy efficiency cannot be considered independently of performance. A configuration that benefits one application may introduce contention, latency or additional runtime for another. Understanding these trade-offs is therefore essential for selecting appropriate execution strategies. The seminar summarizes the main findings from experiments on leadership-class systems and discusses the broader challenges of improving energy efficiency while preserving application performance.

Bio: Matheus M. Costa is a Ph.D. student in computer science at the Federal University of Rio Grande do Sul, where he researches energy-efficient high performance computing (HPC), GPU resource management, in-situ analysis and power-aware execution of coupled HPC and artificial intelligence workflows.

Series: See upcoming and previous presentations at CS Seminar Series.