Cycle-Accurate NoC-based Convolutional Neural Network Simulator

Due to the development of intelligent systems, convolutional neural network (CNN) have been applied and achieved outstanding performance in many aspects, such as patent recognition and object classification. Although CNN brings many advantages to several AI applications, the longer computing time and the larger computing power still restrict the system performance significantly. Therefore, the hardware-efficient CNN accelerator design receives much attention in recent years. However, because of the intensively complicated computation and communication among the CNN operation, the interconnection between each CNN computing unit becomes complicated as the CNN size is scaling up. On the other hand, the Network-on-chip (NoC) interconnection has been proposed to solve the complex communication problem, which is an attractive interconnection to construct the hardware-efficient CNN design. To evaluate the NoC-based CNN design in the system level, we present a cycle-accurate NoC-based convolutional neural network simulator, CNN-Noxim, in this paper. The proposed CNN-Noxim can simulate the CNN models and the classification precision of the simulation output is verified by Keras. Consequently, the proposed NoC-based CNN simulator is a high flexible neural network simulator, which facilitates the evaluation of the NoC-based convolutional neural network design.