Briefly Analysis about CNN Accelerator based on FPGA

Abstract

Since convolutional computation in deep learning is a large and time-consuming computation, researchers often use GPU or FPGA to accelerate these computation. This paper illustrates the advantages of convolutional computation by using FPGA accelerators. In detail, the paper presents some research results of convolutional computation based on FPGA, and explains the current common way of FPGA accelerator design, which uses high-level synthesis and Vitis AI. Finally, we deploy and run YOLOv4 model on the ZCU102 evaluation board using Vitis AI, and perform object detection with the tableware dataset, achieving a recognition result of 96.2%, and 72.5 times higher performance than CPU.

Topics & Concepts