Abstract: Aiming at the contradiction between the huge consumption of computing resources by deep neural network in pursuit of accuracy and the restricted environment of edge computing platforms, a new edge computing fabric is explored. The new fabric uses FPGA logic resources to construct the Neural Network Tensor Processing Units (TPUs), and collaborates with ARM CPU in the FPGA chip, which not only accelerates the computation of deep neural network model and improves the accuracy, but also optimizes the power consumption of the entire system significantly. Under this architecture, the accuracy of the compressed MobileNet-V1 is 78.1%, while the power consumption is only 3.4 W. Its comparison with other deep learning edge computing platforms with different computing fabrics shows that the system has obvious advantages for accelerating the computation of small-scale deep neural networks, without reducing the accuracy.

Key words: FPGA, tensor processing unit, deep learning, edge computing