Energy Efficiency and Timeliness in Model Training for Internet-of-Things Applications: Poster Abstract

Neural network model training is indispensable for domain-specific Artificial Intelligent Internet-of-Things (AIoT) applications. Typically, a GPU graphics card may take several hundreds watts in average during model training, while an embedded GPU device may take only couple watts for the same purpose at the cost of a longer training time. In this paper, we report our empirical study on the model training using NVIDIA RTX 2080 Ti graphics card and NVIDIA Jetson Nano embedded device. We show that, surprisingly, while the training time using the Jetson Nano is 30 times slower than that using the graphics card, the total energy consumption by Jetson Nano is actually only half. The result suggests that when the response time is less critical, one may choose to do model training on GPU embedded devices instead.