Physically Interpretable Power, Capacitance, and Power-Density Modeling for Future GPU Architectures

Modern GPU power models can accurately predict total power consumption, but they often provide limited insight into the underlying physical mechanisms responsible for power dissipation. This internship aims to bridge the gap between performance-counter-based power estimation and physically meaningful architecture characterization. The student will investigate how workload activity, voltage-frequency operating points, and architectural resources contribute to power consumption, and will develop methodologies to separate activity factors from technology-dependent parameters such as effective capacitance. The work will include feature analysis, DVFS-aware modeling, validation using architectural or RTL-level information, and the development of block-level power and power-density models. The ultimate goal is to establish a framework that not only predicts power accurately, but also supports technology-scaling studies, thermal analysis, power-density estimation, and future architecture exploration.