engleski

Determining the Influence of Hardware on the Execution Times of Trained Machine Learning Models

While many discussions and observations have been made regarding the execution times of machine learning (ML) model training, not many researchers have shown concern regarding the execution times of trained models when the model are applied for inference. In this paper, the researchers observe the execution times of a realistic hybrid system consisting of a YOLOv3 based detection model and two classification models based on VGG16 and VGG19 convolutional neural networks. The provided hybrid model is tested on five different hardware configurations single core CPU execution, eight thread CPU execution, 48 thread CPU execution, single GOU execution and five GPU execution. The goal is to compare the execution times and determine the user satisfaction and ease of use on standard consumer hardware, readily available to the end users of the project, in comparison to a high performance computing architectures. The results show that the best performing architecture is a single GPU, but even the slowest solution (single core CPU), does not show an extremely slow time when inference is applied.

rtificial intelligence, convolutional neural networks, execution timing, high performance computing, hybrid systems, machine learning, model inference

nije evidentirano