<p dir="ltr">Smartphones have revolutionized personal computing by blending communication, entertainment, and productivity into one. Emerging applications demand high computational capabilities and are power-hungry. Optimizing the battery drain of mobile apps is critical to extending the battery life of mobile devices, enhancing the user experience. This involves optimizing the power consumption of all power-hungry components in modern smartphones.</p><p dir="ltr">The power draw of a mobile System on a Chip (SoC) is determined by its computational requirements and can become a bottleneck for the entire system. To understand and optimize power consumption, it is essential to accurately characterize the power of heterogeneous SoCs and study the interplay among default governors for the mobile CPU, GPU, and memory. Therefore, power modeling of mobile device components is foundational for effective energy profiling and analysis.</p><p dir="ltr">This thesis presents an understanding of the power consumption on modern smartphones and contributes to creating a display power model for Organic light-emitting diode (OLED), developing a lightweight GPU power model, and finally creating a Unified energy-aware governor for LLM.</p><p dir="ltr">First, we develop a novel piecewise OLED power model that accurately estimates the display power draw as a function of the displayed content. We present the design and implementation of two OLED power management tools: an accurate per-frame OLED display power profiler called PFOP and an enhanced Android Battery called Battery+ that help phone users to understand and manage phone display energy drain.</p><p dir="ltr">Second, we present APGPM, the first mobile GPU power modeling methodology that automatically selects an optimal set of performance monitoring counters (PMCs) that can accurately estimate the GPU power across different workloads. We implement APGPM in Android and evaluate it on two representative mobile GPUs, Qualcomm’s Adreno and ARM’s Mali, for diverse GPU workloads. Our evaluation shows that APGPM builds a GPU power model that reduces the average GPU power prediction error by almost half compared to the prior-art, utilization-frequency based smartphone GPU power model.</p><p dir="ltr">Third, we present a unified energy-aware governor designed to optimize the energy efficiency of three power hungry components, CPU, GPU, and memory, for large language model (LLM) inference on mobile devices. We show that the triplet governors used in mobile OSes such as Android can result in longer prefilling and decoding latencies compared to optimal combinations of CPU/GPU/memory frequencies under the same energy budget. Our unified energy-aware governor, LEG, is shown to significantly reduce time-to-first-token and time-per-output-token of LLM inference compared to the default governors.</p>