Dynamic Voltage and Frequency Scaling (DVFS) Strategy for FPGA-Based Edge AI Inference

Edge Artificial Intelligence (Edge AI) inference accelerators based on Field Programmable Gate Arrays (FPGAs) are increasingly deployed in power-constrained environments, where energy efficiency and low latency are critical. Dynamic Voltage and Frequency Scaling (DVFS) is a well-established technique for power optimization, yet its application in FPGAbased Edge AI systems remains limited. This paper presents a workload-aware DVFS strategy tailored for FPGA-based inference accelerators. The proposed framework integrates a lightweight workload monitoring module and an FSM-based controller to dynamically adjust voltage and frequency according to real-time workload conditions. A Verilog-based simulation on a representative CNN inference pipeline demonstrates that the proposed strategy achieves a 22.2 % reduction in estimated energy consumption without incurring noticeable latency overhead. Additional latency evaluation confirms that real-time responsiveness is preserved across all VF operating modes, highlighting the effectiveness of the proposed approach in balancing energy efficiency and performance for Edge AI deployments.