Mastering power control in HPC CPUs: a journey through modeling, algorithms, and hardware insights

High-Performance Computing (HPC) has rapidly evolved to meet the increasing computational demands of data-intensive fields such as climate modeling, artificial intelligence, and physics research. This growth is driven by the demand for massive computational power that spans various demographics, including researchers, industry professionals, and governments, reaching end-users with the rise of Large Language Models (LLMs). Emerging trends in HPC, including many-core and heterogeneous architectures, present significant complexity, especially as they adopt advanced chiplet-based designs with specialized accelerators. These innovations introduce challenges that necessitate sophisticated control strategies to manage power and thermal dynamics effectively. The open-source RISC-V ISA has spurred the entry of new players into this market segment. However, despite advances in hardware design, there remains a noticeable gap in research concerning on-chip power and thermal control strategies. Existing efforts have largely focused on high-level, software-based control mechanisms at the operating system or application level, leaving low-level control methods underexplored. This thesis addresses this gap by developing and evaluating advanced low-level control algorithms for power and thermal management in HPC environments. It introduces a comprehensive modeling framework that captures essential system dynamics and highlights the unique challenges of low-level control, such as leakage power management, actuator non-idealities, and coupling constraints. The proposed control strategies, including fuzzy-inspired and Model Predictive Control (MPC) approaches, are validated using a Hardware-in-the-Loop (HIL) testing platform to demonstrate their effectiveness in real-time scenarios. Results indicate that these advanced controllers significantly enhance thermal regulation, minimize performance degradation, and achieve superior energy efficiency. The thesis concludes by outlining future research directions, such as integrating machine learning for predictive control and exploring distributed control frameworks to further optimize HPC system performance.