Evaluating the Impact of GPU Frequency Tuning and Power Capping on Performance and Efficiency

SUMMARY High-Performance Computing (HPC) systems play a pivotal role in modern scientific and en- gineering advancements. However, the increasing demand for computational power comes with significant energy consumption challenges. This thesis investigates the impact of GPU fre- quency tuning and power capping on performance and energy efficiency across three generations of NVIDIA GPUs: Pascal (P100), Volta (V100), and Ampere (A100). The study leverages the Altis Benchmark Suite to comprehensively evaluate the performance and energy behavior of diverse workloads. Through systematic power management strategies, including power capping and frequency tuning, we aim to identify optimal configurations that balance computational performance and energy consumption. The results reveal that power capping is particularly effective for compute-bound workloads, while frequency tuning provides finer-grained control over energy efficiency, especially for architecture-specific optimizations. Additionally, the ex- periments demonstrate that the Ampere architecture (A100) offers more precise control over power-performance trade-offs compared to its predecessors, highlighting the evolution of power management capabilities in modern GPUs. A novel contribution of this research is the integra- tion of predictive modeling to estimate energy consumption under varying frequency and power cap configurations. Three machine learning models were employed: Gaussian Process Regres- sion (GPR), Long Short-Term Memory (LSTM) neural networks, and Random Forest (RF). These models were trained and validated on experimental data to forecast energy consump- tion trends, allowing for adaptive energy-aware strategies in HPC environments. The results demonstrate that GPR excels in capturing smooth, non-linear patterns, LSTM effectively han- dles workloads with temporal dependencies, and RF provides robust performance for diverse benchmarks. This predictive framework not only enables real-time optimization but also as- sists in developing energy-efficient scheduling strategies for future HPC systems. Furthermore, this work introduces a classification of benchmark behaviors into consistent and non-consistent categories, aiding in the identification of workloads that benefit the most from specific power management techniques. By establishing clear guidelines for frequency tuning and power cap- ping, this thesis provides practical recommendations for optimizing energy efficiency without compromising performance. In conclusion, this research enhances the understanding of GPU power management techniques and introduces a predictive framework for dynamic energy op- timization in HPC clusters. The findings pave the way for more efficient and sustainable high-performance computing systems, enabling better resource allocation, reduced operational costs, and improved environmental impact