Dynamic Partial Reconfiguration of FPGAs for Energy-Efficient Machine Learning Inference in IoT Systems

This work introduces a dynamic partial reconfiguration (PR) approach for field-programmable gate arrays (FPGAs) to enhance the performance, energy efficiency, and adaptability of IoT systems. By enabling run-time reconfiguration, PR allows flexible allocation of resources for machine learning (ML) tasks, minimizing energy consumption and hardware demands. The method replaces fixed-length floating-point arithmetic with variable bit-width representations, reducing memory usage and computational complexity without sacrificing accuracy. Demonstrated on Graph Convolution Network (GCN)-YOLO models, this approach efficiently handles diverse ML layers while achieving significant improvements in resource utilization, latency, and energy savings. These results establish a scalable framework for real-time edge intelligence in IoT applications.