The energy-intensive nature of modern AI models presents a fundamental challenge to sustainable computing. As these models grow in complexity and deployment, their substantial energy demands during training and inference contribute significantly to carbon emissions and operational costs. This tutorial first examines the technical challenges of measuring energy consumption across different levels of abstraction, from system to process-level measurement and down to individual source code methods and API calls. This part of the tutorial demonstrates practical approaches to overcome these measurement challenges. The next part of the tutorial focuses on power consumption patterns in GPU kernels. We analyze the relationship between power consumption and key kernel characteristics, including thread configuration, block geometry, and power limit settings. We illustrate how these parameters influence energy efficiency and demonstrate approaches for predicting power consumption based on kernel properties. The tutorial concludes with insights into managing the critical performance-energy tradeoffs in GPUaccelerated AI applications.