A Heterogeneous FPGA-based Accelerator Design for Efficient and Low-cost Point Clouds Deep Learning Inference
Jinling Xu, Yonggui Wang, Wenbiao Zhouy · 2022
The neural networks on 3D data and applications have emerged in the past five years. However, there are only a few dedicated hardware designs were proposed for 3D data and algorithms. Meanwhile, they lack flexibility and adaptation for the fast evolvement of software algorithms. We propose a heterogeneous accelerator design on Xilinx Zynq and Zynq UltraScale+ platform. An innovative vector pipeline is designed in the accelerator that can reach the near limitation of BRAM frequency, and it gives the final design frequency closure at 550MHz with 100% DSP usage.