A CNN processing method and device

Abstract

The invention discloses a CNN processing method and device, which are used for realizing the module operation in the CNN model on the ASIC, and reducing the operation cost while ensuring the calculation performance of the CNN. An embodiment of the present invention provides a CNN processing method, including: obtaining the intensive type of the first module in the CNN model; if the intensive type of the first module is computationally intensive, deploying the first module to a dedicated integration on the circuit ASIC, and obtain the multiple computing resources of the ASIC occupied by the first module; merge the same computing resources among the multiple computing resources of the ASIC occupied by the first module, and obtain the deployment The first module on the ASIC after completing the combination of computing resources; running the first module deployed on the ASIC after completing the combining of computing resources.

Description

technical field [0001] The present invention relates to the field of computer technology, in particular to a method and device for processing convolutional neural networks (English full name: Convolutional Neural Networks, English abbreviation: CNN). Background technique [0002] With the continuous development of machine learning, the accuracy of picture recognition is getting higher and higher, and more and more applications have begun to use picture recognition technology, such as the earliest digital recognition, and now it is very widely used face recognition. To make image recognition reach the level of human intelligence, a large number of computing resources must be used for calculations, which has great requirements for computing performance. While meeting the performance requirements, the control of computing costs needs to limit the scale of computing resources consumed. [0003] Currently, the mainstream image recognition algorithm is CNN, which is mainly impleme...

AI Technical Summary

Problems solved by technology

[0005] The biggest advantage of the CNN solution implemented by CPU and GPU is high performance, but currently the cost of deploying GPU is very high, especially the cost of deploying a large GPU cluster is higher, and GPU consumes a lot of power, regardless of hardware deployment cost or follow-up operation cost are very high

Images