Mapping DNNs onto the NoC Platform

Sammanfattning: This thesis uses an existing NoC simulation platform to construct a Network on Chip-based many-core system. The network is an 8_8 mesh topology. This thesis chooses LeNet5, ResNet, VGGNet, and AlexNet as the computing load, and tries to obtain a deep neural network mapping algorithm based on a NoC design method that can be widely used. By analyzing the data structure and operation methods of different deep neural networks, this thesis has discovered some key elements to improve the efficiency of NoC based many-core systems. These elements are: How to divide the different layers of the deep neural network into corresponding tasks? How to simplify the data flow between the tasks of the deep neural network? How to reduce the mutual interference between different data streams? And whether to use parallel execution strategies? In the end, this thesis chose the method of layer merging to reduce the total number of tasks and increase the range of DNN adapted to the NoC platform. The relay transmission mode is used in data transmission to reduce the mutual interference between packets. The depth-first search algorithm is used to reduce the data transmission distance while mapping tasks. The parallel execution mode is added to the platform to improve the throughput of the device when dealing with bursts of large amounts of data. The main technical indicators are the total number of operating cycles, the average transmission distance, and the output frequency during batch processing.