A Lightweight Pipelined Architecture for CNN Deployment on Edge Devices

作者：Ruitong Qiao, Hao Ding, Yi Yang, Yiyun Chen, Jinshan Li, Di Fan, Zi Wang, Yimao Cai · 年份：2026 · DOI：10.1109/cstic68613.2026.11537692 · 研究领域：Innovative Energy Harvesting Technologies、Advanced Memory and Neural Computing、Advanced Neural Network Applications

Multi-core distributed computing is considered a promising approach to address resource constraints in edge convolutional neural network (CNN) deployment. However, existing solutions face two key challenges: high latency and power due to serial operations, and significant hardware overhead from sliding-window convolutions. This paper proposes a lightweight and efficient architecture adopting a channel-first, address-incremental serial dataflow, specially tailored for CNN acceleration on edge devices. Key techniques include: (1) a pipelined design that integrates padding, convolution, and pooling to jointly optimize computation and communication; (2) a lightweight convolution scheme utilizing multiply-accumulate (MAC) operations with low resource usage. Simulations on the Vivado platform show that, in one convolution layer of CNN, the proposed design reduces look-up table (LUT) and flip-flop (FF) usage by 87.3% and 73.1%, respectively, while the overall latency of the pipeline—including padding, convolution, and pooling—is reduced by 10.35%, with negligible accuracy fluctuation below 1%.