Convolutional Neural Network Accelerator Modeling Based On NOC Structure

With the rapid development of artificial intelligence,the core algorithm of artificial intelligence-neural networks has incresingly draw much attention of people.But general-purpose processors can hardly meet the application requirements efficiently.On the layer-by-layer analysis of LeNet algorithm and AlexNet algorithm,this paper designs a convolutional neural network accelerator based on NoC structure according to the single direction transmission characteristics of these two algorithms in the process of data flow.The accelerator uses a ring topology which is superior to the grid structure in terms of bandwidth and latency.In the structure,16 resource nodes are arranged,which are divided into four horizontal rings and four vertical rings.The packet exchange mechanism is adopted to transmit data on the network.The width of the transmission microchip is 130 bits.At simulation frequency of 1 GHz,the theoretical peak bandwidth of the accelerator can reach 2080 Gb/s.A more comprehensive processor core is applied with the resource node to substitute a simple functional processing unit to increase the versatility of the accelerator.SystemC Ianguage is used to establish an accurate clock model.The performance of the entire accelerator network is tested under different injection rates with random stimulator,and the number of received packets and the total delay time of each node are counted.The convolutional neural network is mapped according to the computational volume and inherent algorithm structure of each layer in the LeNet algorithm and the AlexNet algorithm.The divided algorithm is programmed using C program language,which is compiled and linked with the RISC-V cross compiler tool chain.Then,the format is converted and loaded into the accelerator's 16 cores for performance simulation.At 1GHz,the accelerator extracts 10,000 images from the MNIST hand-written digital database.Afterwards,it is processed into 32×32-size data as an input.The speed is improved by 11.86 times compared with single-core.For the AlexNet algorithm,1000 images were randomly extracted from the ImageNet database,which is processed into 224×224×3 data to perform testing.The acceleration effect is 11.03 times to the single-core accelerating effect.