Performance Evaluation Of Deep Learning Frameworks On Embedded Gpu

Deep learning method is more and more widely used in many fields nowadays, bringing great influence and exciting results for researchers. An increasing number of deep learning frameworks are beginning to emerge. As the computation of the convolutional neural network is highly intensive, in order to improve the performance, the developer should carefully choose the framework, especially for the embedded platform with limited resources. In this paper we investigate the performance, power and energy efficiency of three deep learning frameworks (Caffe, MXNet and Darknet) on image classification. The comparison is based on the embedded GPU Tegra TX1, through three different network structures (AlexNet, GoogleNet and VGG-16). Our results show that Caffe has the best performance in the case of single image input, and MXNet gradually shows its advantages in memory optimization and data parallelism. Darknet has the features of concise and less dependent, making it very convenient to achieve the transplantation between different platforms.