Dynamic FPGA reconfiguration for scalable embedded artificial intelligence (AI): A co-design methodology for convolutional neural networks (CNN) acceleration

In recent years, FPGA platforms have shown significant potential for accelerating artificial intelligence (AI) applications, particularly in Embedded AI. While various studies have explored adaptive AI deployment on FPGAs, there remains a gap in methodologies fully integrating software adaptability with FPGA hardware reconfigurability. This article presents a novel end-to-end co-design methodology for deploying adaptable and scalable Convolutional Neural Networks (CNNs) on FPGA platforms. The framework enhances computational performance and reduces latency by dynamically modifying hardware acceleration units by combining CNN architecture adaptability with dynamic partial reconfiguration of FPGA hardware. The proposed methodology enables automated synthesis and runtime customization of both hardware accelerators and CNN architectures, eliminating the need for iterative synthesis. This approach has been implemented and tested on a Xilinx XC7020 FPGA board for a CNN-based image classifier, achieving superior computation performance (0.68s/image) and accuracy (97%) compared to state-of-the-art alternatives. • Adaptive, scalable framework for AI deployment on FPGA via dynamic reconfiguration. • Adaptive CNN-FPGA: Cuts iterative synthesis, speeds deployment, and reduces costs. • Runtime FPGA reconfiguration: Adapts CNNs & hyperparameters without full re-synthesis. • Fine- & coarse-grained hardware customizations improve computation & reduce latency. • Tested on Xilinx ZYBO FPGA: CNN classifier shows improved efficiency & scalability.