Issue |
SHS Web Conf.
Volume 139, 2022
The 4th ETLTC International Conference on ICT Integration in Technical Education (ETLTC2022)
|
|
---|---|---|
Article Number | 03005 | |
Number of page(s) | 8 | |
Section | Topics in Computer Science | |
DOI | https://doi.org/10.1051/shsconf/202213903005 | |
Published online | 13 May 2022 |
Parallelization and Hardware Mapping of Deep Neural Network on Reconfigurable Platform for AI-Enabled Biomedical System
The University of Aizu, School of Computer Science and Engineering, Adaptive Systems Laboratory, Japan
a) Corresponding author: m5251136@u-aizu.ac.jp
b) Electronic mail: ku@ieee.org.
c) Electronic mail: d8222111@u-aizu.ac.jp.
d) Electronic mail: benab@u-aizu.ac.jp.
COVID-19 is still disrupting many parts of the world. A rapid and accurate diagnosis solution is needed to combat the pandemic. As a part of the AIRBiS(AI-Enabled Real-time Pneumonia Detection Bio-medical System), this work conduct hardware acceleration to speed up the diagnosis. We found that more than 90% of the current diagnosis time is spent on the convolution function and have conducted three methods to speed up the convolution operations. Firstly, by applying the Winograd algorithm on convolution layers, the multiplication operations of the matrices can be decreased, which speeds up the calculation. The next step is to improve the data exchange speed between the FPGA and CPU by replacing the normal buffer with LineBuffer. We also tried to improve the calculation speed by quantization, reducing the number of bits used for the filter and the input image. The FPGA board we used for this research is ZCU102. The application used for high-level synthesis is Xilinx SDSoC 2019.1. Using the mentioned approaches, we improved the inference speed from 106ms to 22.2ms per image.
© The Authors, published by EDP Sciences, 2022
This is an Open Access article distributed under the terms of the Creative Commons Attribution License 4.0, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.