Research On Deep Learning Acceleration Technology Based On Domestic Computing Platform

Deep learning,as an essential part of Artificial Intelligence,has been wildly applied in the industries.Under the circumstances,the domestication of computing devices is increasingly important.However,domestic computing platforms lack of available computation accelerators,which makes it hard to generalize.In this article,we found two ways to realize acceleration of deep learning on domestic computing platform,including the technique of supporting GPU and technique of the optimization of domestic FPGA.Our work can be divided into three points,firstly we referred to the literature to make certain problems of the acceleration devices on domestic computing platform,and determined the research direction of two acceleration techniques.Secondly,we researched some framework of common-used GPU computing platforms,then selected appropriate GPU for domestic computing platform.Some technical approaches including cross-compilation,kernel module replacement and system environment variable setting were applied for building the GPU general computing platform which make GPU can be used on domestic computing platform.Thirdly,for the contradiction between insufficient domestic FPGA logic resources and the requirement from deep learning,we optimized the convolutional neural network by matrix compression and data quantization,then realized the acceleration for VGG through domestic FPGA.The result of experiments shows that the technique in this article can successfully accelerate the computing of deep learning.Compared with domestic CPU,the computing efficiency of GPU and domestic FPGA can be improved by 48 times and 284 times respectively,and promoted application of deep learning on domestic computing platforms wildly.