Task-Aware and Predictive Event-Driven Power Gating Architecture for Energy-Efficient Edge AI Hardware Systems

Background: The rapid integration of multi-modular Artificial Intelligence (AI) components into edge devices has led to asignificant increase in static and dynamic power consumption. Conventional power management strategies often lack thegranularity required to address the deterministic and sequential nature of AI task execution, resulting in substantial energy leakage.Objective: This study proposes an enhanced Task-Aware Event-Driven Power Gating (TA-EDPG) architecture. The frameworkaims to synchronize hardware power states with chronological task requirements while addressing the critical challenges oftransition latency, hardware reliability, and complex workload scalability.Methodology: The proposed architecture implements a Deterministic Power Orchestrator augmented with three advanced layers:Adaptive Predictive Layer (APL): Utilizing first-order Markov Chain models to anticipate task transitions and preemptivelyinitiate module warm-up, effectively neutralizing gating-induced delays.DAG-based Scalability: Generalizing the orchestration logic to support Directed Acyclic Graph (DAG)taskflows, enablingenergy-efficient management of parallel and multimodal AI pipelines.Reliability Modeling: Incorporating a Coffin-Manson-based thermal stress analysis to quantify the impact of rapid power-cyclingon MOSFET longevity, ensuring hardware sustainability. Validation was performed via MATLAB/Simulink and empirical testingon an ARM-based platform.Results: Experimental results demonstrate that the TA-EDPG architecture achieves up to 93.6% energy savings during standbyand 37.1% during active operation, while the Adaptive Predictive Layer reduces wake-up latency by over 95% (from 150 ms to6.8 ms) without compromising energy efficiency.Conclusion: This research establishes a comprehensive, intelligent, and hardware-sustainable power management paradigm,providing a scalable solution for next-generation, battery-constrained Edge AI platforms.