Design of 2D Systolic Array Accelerator for Quantized Convolutional Neural Networks

Chia Ning Liu, Yu An Lai, Chih Hung Kuo, Shi An Zhan

Research output: Chapter in Book/Report/Conference proceedingConference contribution

2 Citations (Scopus)

Abstract

Quantization techniques have been studied to reduce the computing and memory requirement of deep neural networks. The full precision floating-point numbers are quantized into integer representation with lower bit-width. In this work, we quantize both activations and weights in CNN to 8-bit integers and apply our quantization method to the hardware accelerator. The accelerator is designed with a systolic-based structure, which can support both the convolutional layers and the fully-connected layers for various network models. By the proposed quantization scheme, there is only 1.68% mAP loss on YOLOv3-tiny model compared to the floating-point model. Benchmarked with AlexNet and VGG-16, the external memory access of convolutional layers is reduced by 1.63x and 1.79x compared with Eyeriss, and the internal memory access is also reduced by 7.31x and 17.48x.

Original languageEnglish
Title of host publication2021 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2021 - Proceedings
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781665419154
DOIs
Publication statusPublished - 2021 Apr 19
Event2021 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2021 - Hsinchu, Taiwan
Duration: 2021 Apr 192021 Apr 22

Publication series

Name2021 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2021 - Proceedings

Conference

Conference2021 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2021
Country/TerritoryTaiwan
CityHsinchu
Period21-04-1921-04-22

All Science Journal Classification (ASJC) codes

  • Hardware and Architecture
  • Safety, Risk, Reliability and Quality
  • Instrumentation
  • Electrical and Electronic Engineering

Fingerprint

Dive into the research topics of 'Design of 2D Systolic Array Accelerator for Quantized Convolutional Neural Networks'. Together they form a unique fingerprint.

Cite this