Convolutional Neural Network Accelerator for Integrated Winograd and Convolution Computations

Due to the development of modern artificial intelligence, CNN has been widely used in many application domains. Because the architecture of CNNs has become more and more complicated in recent years, the amount of computation time increases dramatically. In order to meet the need of real-time computi...

Full description

Saved in:
Bibliographic Details
Main Authors: Lo, Yueh-Fan, Cheng, Wei-Kai
Format: Conference Proceeding
Language:English
Subjects:
Benchmark testing
CNN
Computer architecture
Convolution
Convolutional neural networks
Kernel
PE architecture
Real-time systems
systolic array
Systolic arrays
Winograd
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Due to the development of modern artificial intelligence, CNN has been widely used in many application domains. Because the architecture of CNNs has become more and more complicated in recent years, the amount of computation time increases dramatically. In order to meet the need of real-time computing, some research use the Winograd algorithm to reduce the number of multiplications. However, Winograd algorithm does not get computation benefit for all kernel size and stride. In this paper, we propose an integrated PE (processing element) architecture that can support both Winograd and traditional convolution computations. Winograd architecture is used for the most common 3x3 kernel with stride-1 sliding window calculation, and convolution operation is performed on the systolic array architecture for the other kernel size or non-stride-1 sliding window. Based on the proposed integrated Winograd and convolution PE architecture, we can reduce quite amount of the CNN computation cycles.
ISSN:2575-8284
DOI:10.1109/ICCE-Taiwan58799.2023.10226753