LRSA: A Low-Latency Reconfigurable Systolic Array for Kernel-Dynamic CNNs on Edge Devices

View at publisher Summarise with VIRAT
Authors: Zih-Jia Wang, Shu-Yen Lin
Cited by: 0
References: 11

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study presents a Low-Latency Reconfigurable Systolic Array (LRSA) architecture designed to enhance the computational efficiency of CNNs in edge-AI platforms. Unlike conventional systolic arrays that suffer from increased latency and degraded performance caused by inefficient PE utilization and bypass delays, LRSA introduces a reconfigurable PE interconnection supporting multiple dataflows and variable sub-array widths for workload adaptation. Experimental results show that LRSA achieves 0.58× lower latency and 3% higher PE utilization compared with the related work, the lightweight architecture for supporting fine-grained reshaping and multiple dataflows on systolic array (ReDas). Overall power consumption decreases by 15.7%, demonstrating an improved balance between performance and energy efficiency. With its low-latency operation, power-efficient design, and flexible configurability, LRSA enhanced suitability for real-time AI acceleration in edge and consumer electronic devices.

Fields of Study & Subjects

Advanced Neural Network Applications Parallel Computing and Optimization Techniques Embedded Systems Design Techniques

Publication details

Journal / source
International Conference on Consumer Electronics
Pages
1–6
Publication date
3 February 2026
Language
EN