Cambricon provides us with a complete intelligent application system, how to use this system for deep learning algorithms development is a challenging issue. In this paper, we exploit, evaluate and validate the performance of the ResNet101 image classification network on Cambricon with Cambricon Caffe framework, demonstrating the availability and ease of use of this system. Experiments with various operational modes and the processes of model inference show, the optimal running time of a common ResNet101 network that classifies the CIFAR-10 dataset on Cambricon is nearly three times faster than the baseline. We hope that this work will provide a simple baseline for further exploration of the performance of convolutional neural network on Cambricon.
Supplementary notes can be added here, including code and math.