Performance Evaluation of CPUs and Edge GPUs on Time-Series ML Applications

This study evaluates the trainability and performance of a deep learning model working on time-series data in an edge hardware environment. LSTM architecture was used as the model, and univariate predictions were performed with the UCI Electricity Load Diagrams dataset, consisting of four years of electricity consumption data. The aim is to examine whether not only inference but also the direct training process can be implemented on embedded hardware. The experiments were conducted on both a computer with valid processing power and an NVIDIA Jetson AGX Orin device; single and parallel training scenarios were performed with different amounts of house data (1, 3, 10, and 50). Criteria such as training time and evaluation metrics were analyzed comparatively. Despite some architectural limitations on the Jetson device, it was observed that the model could be trained successfully, the obtained prediction results offered evaluation metric levels close to those of the computer environment, and it has been proven that edge devices can be used as alternatives in the field. This study also demonstrates that edge hardware can be a practical option not only for low-latency inference but also for operations such as model updating or retraining in the industry. The findings aim to contribute to the more widespread and effective use of edge AI applications in resource-sensitive sectors such as energy.