An Acceleration Structure Of Convolutional Neural Network

Artificial intelligence technology has gradually become mature and practical with the help of the Internet,big data,high-performance chips and other technologies.With the continuous improvement of basic technologies such as DNN,artificial intelligence will enter various industries and radically change people's production and lifestyle.CNN is the representative algorithm of DNN,and cause extensive attention of highly efficient classification recognition in recent years.Its special structure with local weights sharing has unique advantages to reduce the complexity of the networks,and its layout is closer to real biological neural network;especially multidimensional networks input vector image can directly input this feature to avoid the extraction and classification.CNN algorithm has been widely used in machine learning,speech recognition,,document analysis,language detection and image recognition of artificial intelligence.The cloud needs to process massive data with high accuracy and needs certain universality.Due to the limitations of security,real-time and network bandwidth resources,CNN sinks to the terminal.It is of great research significance to study the CNN acceleration structure,whether to accelerate the cloud training speed,improve the calculation power and power ratio,or to meet the needs of terminal to improve the calculation power and reduce the cost and power consumption.Studies show that more than 90% of CNN's computation and time consumption are concentrated in the convolution layer,involving a large number of multiplication and addition operations.The systolic array represents a PE network that rhythmically computing and transmitting data through the system,regularly pumping in and out to maintain a regular flow of data.Making the data flow in computing units,it reduces the memory access,and makes the structure more orderly,more uniform wiring,and higher frequency.Systolic array is a special design with simple structure and low implementation cost.The CNN convolution operation is an ideal application to show the characteristics of systolic array.Therefore,this thesis proposes an efficient array architecture for CNN convolutional layers based on systolic arrays,which accelerates the convolution operation by changing the data flow and computing structure.And further improved on it and proposed an acceleration architecture for CNN convolutional layer based on broadcast architecture,which has greatly improved its throughput and power efficiency.