P‐3.10: Real‐Time YOLOv3‐Tiny Acceleration on Zynq‐7035 Using a Hardware‐Software Co‐Design Approach
SID Symposium Digest of Technical Papers · 2026-08 · 57 (S1) · 1121-1123
DOI: 10.1002/sdtp.70003_302 ↗Kurz gefasst
Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Hanyang Ye und 6 weiteren Autor:innen. Die Arbeit nennt 7 Literaturangaben.
Abstract (Originalsprache)
FPGA system‐on‐chip (SoC) platforms play an important role in edge computing. However, deploying convolutional neural networks such as YOLOv3‐tiny requires an efficient hardware and software co‐design strategy. Automated deployment frameworks, including Vitis AI, significantly reduce development time but often generate opaque implementations that limit architectural transparency and fine‐grained optimization. To address these limitations, this paper presents a manual registertransfer level (RTL) methodology for deploying YOLOv3‐tiny on the Zynq 7035 platform using Vivado 2023.1 and Vitis 2023.1. The core contribution is a deep processing system (PS) and programmable logic (PL) co‐design. On the PL side, a custom convolutional neural network (CNN) accelerator is developed with three key optimizations. First, a pipelined two‐by‐two MaxPool module is designed using a three‐stage two‐by‐one comparison scheme to reduce latency. Second, INT8 quantization is adopted to reduce external memory bandwidth and efficiently exploit the native DSP48E1 blocks. Third, batch normalization parameters are fused into convolution layers to simplify the inference datapath. On the PS side, a C / C ++ application manages task scheduling, DMA control, and post‐processing, including a configurable non‐maximum suppression module with a confidence threshold to enhance detection stability. Implemented on a Zynq 7035 device, the proposed system achieves a throughput of 30 frames per second at 200 MHz while utilizing 18.9 K LUTs, 147 BRAMs, and 304 DSPs, corresponding to approximately 30 percent of the available hardware resources. Experimental results demonstrate real‐time object detection capability. This work provides a transparent and controllable alternative to HLSbased deployment approaches and offers valuable insights for FPGA‐based CNN acceleration requiring fine‐grained hardware optimization.
Zitieren
APA
Ye, H., Wen, X., Ying, X., Sun, R., Ding, S., Zhang, S., & Tang, Y. (2026). P‐3.10: Real‐Time YOLOv3‐Tiny Acceleration on Zynq‐7035 Using a Hardware‐Software Co‐Design Approach. SID Symposium Digest of Technical Papers, 57(S1), 1121-1123. https://doi.org/10.1002/sdtp.70003_302
Vancouver
Ye H, Wen X, Ying X, Sun R, Ding S, Zhang S, et al. P‐3.10: Real‐Time YOLOv3‐Tiny Acceleration on Zynq‐7035 Using a Hardware‐Software Co‐Design Approach. SID Symposium Digest of Technical Papers. 2026;57(S1):1121-1123. doi:10.1002/sdtp.70003_302
BibTeX
@article{Ye2026,
author = {Ye, Hanyang and Wen, Xiaoxuan and Ying, Xiaoyu and Sun, Rui and Ding, Shihao and Zhang, Shengli and Tang, Ying},
title = {P‐3.10: Real‐Time YOLOv3‐Tiny Acceleration on Zynq‐7035 Using a Hardware‐Software Co‐Design Approach},
journal = {SID Symposium Digest of Technical Papers},
year = {2026},
volume = {57},
number = {S1},
pages = {1121--1123},
doi = {10.1002/sdtp.70003_302}
}
EUVIMED