Аннотация
Deep learning models such as convolutional neural networks (CNNs) have a wide range of perception applications in image classification and object detection. However, despite the same CNN architectures, inference performance is different from implementations of specific libraries such as cuBLAS, cuDNN, and TensorRT. To investigate the performance effects of the state-of-the-art GPU libraries, this paper performs a case study of comparison and performance analysis of cuBLAS, cuDNN, and TensorRT implementations/libraries on YOLOv4-tiny, introducing crucial nvprof metrics for fair comparison and rationales of different performance and proposing interpretable machine learning (ML) model-based analysis. The results of our interpretable ML models show 100% accuracy in the classification and 0.0094 MAPE in the regression tasks respectively.