Project Description
The research question is to analyze the impact of hardware and algorithm configurations on the efficiency of the deep neural network training process, from the perspectives of power consumption and response latency. The research activities will be based on embedded devices of NVIDIA Jetson Nano, which have common resource constraints in applications that leverage AI for decision-making and smart control, providing strong motivation for power consumption optimization. Jetson Nano is free from most background processes in computers when profiling energy usage during AI model training, such that interference from other processes can be minimized. To achieve these objectives, first, we look for reliable ways to measure the power usage for the whole board and for single layers of a deep neural network by reviewing the state-of-the-art testing methods. The potential methods include measuring the power of existing sensors or using available tools on the board and separating the power usage of background processes from the AI training process. Then, we vary the hardware and software configuration to systematically collect the power profiles for training neural network models. Finally, the relationship between power consumption and training configurations will be analyzed and formulated as an optimization problem.
Tasks and Responsibilites
The student's tasks and responsibilities include a literature review by conducting thorough research to understand existing methodologies and findings related to power consumption in AI training. This involves reading academic papers, articles, and other relevant materials to grasp the current state of the field. In addition, he is responsible for assisting in setting up experiments to measure power consumption during AI training sessions. This could involve working with specialized equipment for power measurement and ensuring data accuracy. Furthermore, the student will analyze the collected data using statistical tools and software, whose role might include organizing data, performing calculations, and generating visualizations to interpret the results effectively.