Research On Deep Learning Inference Chips Based On NVDLA

With the development of artificial intelligence technology,machine learning methods and neural network structure are becoming more and more complex,and huge data calculations make high requirements for computing chip performance to deal with such applications.Especially the emergence of deep learning technology,the amount of neural network layer structure and internal processing data is getting larger,which requires the processing chip has good adaptability and performance.There are more and more application scenarios of deep learning and neural networks nowadays.According to the task quantity and the application scene characteristics,high-efficiency chip solutions become very important in deep learning and neural network applications with considering the cost of computing chip.The main chip solutions for artificial intelligence applications include general-purpose chip processing solution such as CPU and GPU,FPGA chip processing solution which constitute heterogeneous computing and application specific integrated circuit(ASIC)solution at present.Each solution has its advantages and disadvantages,and should be arranged according to the scene.In the exploration of deep learning and neural network chip solutions,the main processor with the acceleration module to support deep learning and neural network applications provides a new idea.The computing chips of this solution can be called AI chips,with corresponding computational advantages and scene adaptability.Related chips which constitute an AI processor have high computational efficiency in mobile devices and lightweight applications.This thesis researches on the deep learning acceleration module of AI chip solution,especially based on NVIDIA deep learning accelerator(NVDLA).The thesis expounds the related knowledge of deep learning and neural networks,including the overview of machine learning methods,the basic structure of neural networks and two types of typical deep learning neural networks which contain deep neural network and convolutional neural network.The research focuses on each stage of the process in the inference operation of convolutional neural network,and analyzes the NVDLA internal chip modules.The thesis researches on hardware architectural specification,software environment and virtual platform of NVDLA and the research validates the basic functions and the supporting for deep learning frameworks and neural network structure of NVDLA using two classical convolutional neural network models as test cases,which contain LeNet and AlexNet based on Caffe deep learning framework.At the same time,the research is supported by UMC 80nm process.Chip logic synthesis and physics implementation for part of the NVDLA core modules which include the single data point processor,the planar data processor and the cross-channel data processor of NVDLA are completed by the 80nm process of UMC in this thesis,and the relevant NVDLA hardware design parameters are introduced which affect the chip scale.The research has completed the exploratory validation and evaluation of NVDLA,with guiding for the integration and application of NVDLA.The analysis of NVDLA internal core modules,convolutional neural network feature structure and NVDLA internal dataflow has certain reference significance to the designing of deep learning accelerator hardware modules and related software scheduling schemes.