The power consumption of embedded GPU computers for deep learning applications

Abstract

The widespread use of GPU processors has significantly enhanced the usage of various artificial intelligence algorithms, especially machine and deep learning models. However, more capabilities introduced significant challenges in energy consumption, particularly on autonomous platforms with battery and autonomy constraints. Our research is crucial as it addresses the energy consumption issue on desktop and embedded GPU computers from a new angle, providing a practical and optimal measurement solution. We engaged various deep learning models to analyze and determine how different datasets for training and problems for inference affect energy consumption. Two separate platforms were used for the experiments: the NVIDIA Jetson NANO platform, a well-known type of embedded GPU computer widely used in deep learning applications, for the inference process just, and a desktop computer with two GeForce RTX 2060 for the training and inference processes. The image classification problem was involved during the training process on two separate datasets. In contrast, inference problems for object detection and drone classification were engaged on live videos and recorded RF signals during the inference process. Our methodology involved a systematic comparison of energy consumption across different models and datasets, ensuring the validity and reliability of our findings. Our findings underscore the potential to reduce energy consumption on embedded GPU computers by implementing a suitable deep learning model. This not only preserves the performance of the required process (drone radio frequency signal detection and identification and object detection in the video stream) but also paves the way for their effective use in real-life scenarios, thereby addressing a crucial need in deep learning. More importantly, we created an empirical methodology for measuring the energy consumption of embedded GPU computers for deep learning applications, directly impacting the development of energy-efficient deep learning systems.