Development And Algorithm Research Of AI Chip Software Stack

With Moore's law coming to an end,it will be more and more difficult to compete hardware performance by improving technology.Under the current technology,how to improve the hardware architecture to adapt to the changing AI algorithm becomes more and more important.In order to study the design of the interface from the bottom hardware to the top software for specific applications,and ensure that the design of the interface between the bottom and the top software meets the requirements of specific applications,this paper develops the required software stack based on the AI chip research platform based on the open-source hardware framework and the FPGA open-source software framework PYNQ,and verifies the software stack through the improved YOLOv3-Tiny object detection algorithm.The main work includes:(1)The overall framework of AI chip software stack is designed,and the software development of AI chip research platform is completed,including model training,model transformation,CNN acceleration instruction and hardware driver development,and CNN application level API call interface development.(2)Transplant and test the object detection algorithm of specific application scenario in the platform.There are some problems in the application of traditional detection methods in pig detection and counting,such as vulnerable detection equipment and unsatisfactory detection results.In this paper,based on the YOLOv3-Tiny algorithm,a detection scale is added and feature dense connection multiplexing is carried out to improve the accuracy of object detection algorithm.Finally,YOLOv3-Tiny algorithm was transplanted to the artificial intelligence chip research platform for testing and comparison,the test results show that the design scheme has achieved good results.