Convolutional neural network inference accelerator and method

Abstract

The invention discloses a convolutional neural network inference accelerator and method. The accelerator comprises an input image buffer module which comprises N buffers for loading input image data,and N*N operation units which are connected to an input image buffer module and are used for performing convolution operations, wherein each of the buffers stores data of one row corresponding to an image, the N*N operation units support a pulsating form of the image data transmitted between adjacent operation units and are connected to the operation units in the buffers to read image data from the buffer, and remaining operation units read the image data from an adjacent operation unit. According to the convolutional neural network inference accelerator and the method, a bidirectional pulsation array is designed for data reusability brought by a convolutional neural network, the loading efficiency of the data is improved, and thus the convolutional neural network is accelerated.

Description

technical field [0001] The present invention relates to a dedicated acceleration framework for convolutional neural networks, in particular to a convolutional neural network reasoning accelerator and method based on bidirectional pulsation and multi-stage pipelines for accelerating its reasoning operation speed in the reasoning stage of convolutional neural networks . Background technique [0002] Convolutional neural network is a kind of feedforward neural network, which is often used in image recognition, and generally includes convolutional layer, pooling layer and fully connected layer. The convolution operation of the convolution layer is to multiply each weight in the convolution kernel with its corresponding input data point-to-point, and then accumulate the point multiplication results to obtain an output data, and then set according to the step size of the convolution layer Set, slide the convolution kernel, and repeat the above operation. [0003] At present, the...

AI Technical Summary

Problems solved by technology

Song L et al. "A Pipelined ReRAM-Based Accelerator for Deep Learning" (IEEE International Symposium on High PERFORMANCE Computer Architecture.IEEE, 2017:541-552.) also used the characteristics of memristors to realize the convolutional neural network. Backpropagation and backpropagation provide future generations with a new idea of ​​accelerator design, but there is a problem with the use of new materials for accelerator design, that is, the real performance of new materials cannot be considered because they have not yet been put into the market. Temporarily unable to be applied in actual development

Images