2-Bit Quantizated CNN Accelerator on FPGA

  • Boyu Zhu*
  • , Yanjun Zhang
  • , Xu Yang
  • , Gongyu Fan
  • , Jingyi Chen
  • , Yuteng Cao
  • , Chongchang Xu
  • , Yongxiang Zhou
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

In recent years, neural networks have seen extensive applications across various fields due to their high accuracy and ease of use. However, the computational demands of neural networks are extraordinarily high, making it challenging to meet real-time requirements in resource-constrained scenarios. To address this issue, this paper applies a neural network training algorithm that quantizes the neural network to a 2-bit precision without significantly affecting its accuracy. To further improve the real-time performance of the neural network, a dedicated computational unit is designed for it. The proposed neural network accelerator includes a learnable threshold non-uniform quantization module, a 2-bit neural network multiplier unit, and a convolution quantization fusion module. On the Xilinx MPSoC ZU2EG chip, the accelerator achieves a peak performance of 512 GOPS while utilizing only 18,597 LUTs.

Original languageEnglish
Title of host publication2025 5th International Conference on Neural Networks, Information and Communication Engineering, NNICE 2025
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1297-1300
Number of pages4
ISBN (Electronic)9798331507961
DOIs
Publication statusPublished - 2025
Externally publishedYes
Event5th International Conference on Neural Networks, Information and Communication Engineering, NNICE 2025 - Guangzhou, China
Duration: 10 Jan 202512 Jan 2025

Publication series

Name2025 5th International Conference on Neural Networks, Information and Communication Engineering, NNICE 2025

Conference

Conference5th International Conference on Neural Networks, Information and Communication Engineering, NNICE 2025
Country/TerritoryChina
CityGuangzhou
Period10/01/2512/01/25

Keywords

  • component
  • FPGA implementation
  • neural network
  • quantization

Fingerprint

Dive into the research topics of '2-Bit Quantizated CNN Accelerator on FPGA'. Together they form a unique fingerprint.

Cite this