Adaptive Machine Learning-based Proactive Thermal Management for NoC Systems

Due to the intricate interconnections in modern multi-core systems, Network-on-Chip (NoC) technology has proven to be an effective solution for addressing communication challenges. However, the primary design obstacle in current NoC systems is the thermal issue, resulting from diverse workload distribution and high power density. To address this, Proactive Dynamic Thermal Management (PDTM) is implemented to efficiently regulate system temperature by utilizing predicted temperature data to minimize performance impact during temperature control periods. Nonetheless, existing temperature prediction models, reliant on specific physical parameters, often produce significant prediction errors due to temperature sensitivity. In this talk, a novel adaptive machine learning (ML)-based PDTM was introduced, receiving the IEEE TVLSI Best Paper Award in 2024. The adaptive ML-based PDTM initially employs an Adaptive Single Layer Perceptron (ASLP), incorporating a single-neuron operation and LMS adaptive filter technology, to accurately forecast future temperature. Subsequently, adaptive reinforcement learning is employed to determine the appropriate throttling ratio for temperature control. This approach allows the adaptive ML-based PDTM to adjust to the temperature behavior of the NoC system and provide a suitable temperature control strategy in real time.