Thermal-Aware Real-Time Scheduling for Heterogeneous Multicore Edge AI Systems

Abstract

Currently, real-time edge AI increasingly relies on heterogeneous multi-core processors. While the performance asymmetry of these processors can improve computational efficiency, it also increases the complexity of latency, energy consumption, and thermal management. This paper proposes a thermally-aware scheduling framework that coordinates dynamic task prioritization, task-core affinity, short-term temperature prediction, and dynamic voltage and frequency regulation. Since temperature prediction can become unreliable under bursty workloads, this framework combines adaptive decision-making with explicit deadlines and temperature protection mechanisms, rather than relying on unconstrained learning strategies. The proposed evaluation method considers average and tail latency, deadline fulfillment rate, energy consumption of each timely completed task, peak temperature, throttling risk, and scheduling overhead. Analysis shows that thermal states can be viewed as a dynamically scheduled resource, but actual performance improvements still depend on workload composition, environmental conditions, sensor latency, and platform-specific thermal coupling characteristics. Further research is needed to examine the impact of processor aging and cross-accelerator deployment.