CNN processing device based on memristor memory calculation and working method thereof

Abstract

The invention discloses a CNN processing device based on memristor memory calculation and a working method thereof. The method belongs to the field of nonvolatile memories and neural networks. Each crossbar switch matrix of the ReRAM is divided into a memory and a cache; aLU is added, each bit line is connected with an adjacent bit line on the lower side through a transmission gate; to-be-processed input data is stored in a memory; sequentially inputting the weights of the convolution kernels into the word lines according to columns; storing the data on the bit line into a cache; accumulatingthe result to obtain a convolution result so as to obtain convolution layer output, completing the calculation of the excitation layer and the pooling layer through the ALU, repeating the steps untilthe input of the full connection layer is obtained, storing the input into the first column of the processing unit array memory, sequentially inputting each group of weights into corresponding word lines in the memory, and obtaining an output result from the bit lines. According to the invention, the memory capacity of the CNN processing device can be improved, the expenditure of reading input data is reduced, the degree of parallelism of operation is improved, and the speed of CNN processing is improved.

Description

technical field [0001] The invention belongs to the field of nonvolatile storage and neural network, and in particular relates to a CNN processing device based on memristor memory computing and a working method thereof. Background technique [0002] At present, neural networks such as Convolutional Neural Network (CNN) are very common methods for performing tasks such as target recognition, image detection and segmentation, and input data such as images are usually non-negative values. However, the implementation of convolutional neural networks involves a large number of operations, especially convolution operations. Therefore, it takes a very high time and power consumption to implement CNN through traditional CPUs and other processors with low parallelism. Implementing CNN through a processor with high parallelism such as GPU, because it does not have a special design for CNN itself, will also result in low resource utilization efficiency. [0003] In recent years, dedic...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the shortcomings of insufficient memory capacity, serious memory-wall problems, and low parallelism in operations in existing neural network processors, the present invention proposes a CNN processing device based on memristor memory computing and its working method , the neural network-oriented in-memory computing (PIM) architecture has been improved, and the processing array can provide higher memory capacity and higher parallelism when facing large neuron data such as high-resolution images. Neural networks such as processing high-resolution images and other tasks have a very high efficiency improvement

Images