3D convolution operation device and method based on three-dimensional phase change memory

Abstract

The invention discloses a 3D convolution operation device and a method based on a three-dimensional phase change memory. The 3D convolution operation device comprises the three-dimensional phase change memory, an input control module, a state setting module and an output control module. 3D convolution operation is performed by adopting the three-dimensional phase change memory, and phase change units on the same bit line form a phase change unit array corresponding to a convolution kernel; the invention discloses a multi-layer stacked structure based on a three-dimensional phase change memory.An upper electrode and a lower electrode of the three-dimensional phase change memory are used as information input ends, convolution is carried out after the information passes through respective phase change unit arrays, obtained convolution results are superposed on a middle electrode in a current form, the sum of convolution calculation results of input information of the upper electrode andthe lower electrode is obtained, and 3D convolution operation is completed in one step. The operation speed is high, and the operation power consumption is low. Besides, the three-dimensional phase change memory comprises a memory array with a three-dimensional structure, so that the occupied area is small, the integration level is high, and the 3D convolution operation speed can be greatly increased under the condition of smaller occupied area.

Description

technical field [0001] The invention belongs to the field of microelectronic devices, and more specifically relates to a 3D convolution operation device and method based on a three-dimensional phase-change memory. Background technique [0002] Convolutional Neural Network (CNN) is a feed-forward neural network that includes convolution operations. It includes convolutional layers, pooling layers, and fully connected layers. The data is input from the convolutional layer. In the convolutional layer, the input data is convoluted through the convolution kernel to extract the feature value, and then the number of data and parameters is compressed through the pooling layer. After several convolutions and pooling, The data is passed through the fully connected layer and used as a classifier. Convolutional neural network has a unique advantage in pattern recognition with its special structure of local weight sharing. Weight sharing reduces the complexity of the network and is wid...

AI Technical Summary

Problems solved by technology

[0007] Aiming at the defects of the prior art, the purpose of the present invention is to propose a 3D convolution operation device and method based on a three-dimensional phase change memory, aiming at solving the problem that the prior art requires two adjacent input The problem of slow operation speed caused by the convolution operation of the data and then the addition and step-by-step operation of the integration module

Images