Neural network data serial pipeline processing method facing artificial intelligence calculation

Abstract

The present invention relates to a neural network data serial pipeline processing method facing artificial intelligence calculation. The method comprises the steps of: after a layer data processing module at a first layer of a neural network receives initial data, performing in-layer parallel operation; after a layer data processing module at a middle layer of the neural network receives an operation result being subjected to serial output by a last layer, performing in-layer parallel operation; and finally, after a layer data processing module of a final layer receives an operation result being subjected to serial output by a laser layer, performing in-layer parallel operation, and performing serial output of an operation result, wherein if the initial data is input for many times, all the layer data processing modules are employed to perform pipeline processing of the initial data input for many times. The neural network data serial pipeline processing method allow the layers of theneural network to respectively respond to different layer data processing modules, and each layer data processing module performs in-layer parallel operation for this layer of data without external interaction so as to reduce the problems that a lot of power loss and delay are generated caused by continually storing read-write data and parameters from the outside, and the multi-time input data canbe subjected to pipeline processing to improve the processing efficiency of the data of the neural network.

Description

technical field [0001] The embodiment of the present invention relates to the field of computer technology, and in particular to a neural network data serial pipeline processing method oriented to artificial intelligence calculation. Background technique [0002] Neural network is the most widely used tool in the field of artificial intelligence. There are many types of neural networks. Taking deep convolutional neural network as an example, different models of deep convolutional neural networks have different numbers of layers, but there are six main types of calculations. There are three methods: full connection, convolution, pooling, nonlinearity, vector operation and matrix addition. These six methods are mature technologies and will not be described in this article. [0003] The calculation method of the neural network model is essentially the multiplication and addition operation of small matrices. The huge parallelism brings a huge amount of data, and the continuous r...

AI Technical Summary

Problems solved by technology

[0003] The calculation method of the neural network model is essentially a multiplication and addition operation of small matrices. The huge parallelism brings a huge amount of data, and the continuous reading and writing of data and parameters from external storage is its biggest bottleneck, which will generate a lot of power consumption and delay.

Images