Estudo comparativo do desempenho de bibliotecas para redes neurais convolucionais em diferentes microarquiteturas de GPU

Abstract

Context: The popularization of Deep Learning and the Deep Inference spurred the development of tools for the implementation of Convolutional Neural Networks (CNNs) as specific libraries for the development of CNNs, the popularization also stimulated the development of GPUs with CNN acceleration capabilities. Developers and scientists working with CNNs need experimental scientific studies that point out which library is most suitable for a particular GPU microarchitecture. Objective: Compare the performance of the CNTK, PyTorch, TensorFlow1.15 and TensorFlow 2.2 libraries in different GPU microarchitectures (Kepler, Maxwell, Pascal and Turing) using LeNet-5 CNN and the MNIST dataset. Method: Initially, a systematic mapping was carried out as a way to identify and systematize the main benchmarks, then an experiment was carried out evaluating the impact of the Keras API on the performance of the libraries used in the comparative study and lastly, four experiments were carried out comparing the libraries, each experiment was performed in a computational environment accelerated by a GPU with a different microarchitecture. Results: For the systematic mapping, it was identified that the classic CNN architectures such as LeNet-5 and AlexNet are the most used as benchmarks, also showed that the most used data benchmarking sets are ImageNet, MNIST and CIFAR-10. The experiment on the impact of the Keras API showed that the API negatively impacts the performance of all the tested libraries. In the comparative study of the libraries’ performance, the PyTorch library presented the performance performance and the CNTK, TensorFlow 1.15 and TensorFlow 2.2 libraries alternated between the three shorter execution times. Conclusion: The results show that the PyTorch library presents a low level of use of GPU and uses a large amount of sec memory compared to other libraries that showed a higher performance.