Optimising thermal sensor placement and thermal maps reconstruction for microprocessors using simulated annealing algorithm based on PCA

Using embedded thermal sensors, high-performance microprocessors employ dynamic thermal management techniques to measure runtime thermal behaviour so as to prevent thermal runaway situations. However, on-chip thermal sensors are highly susceptible to noise, which results in a higher probability of false alarms and unnecessary responses. In this study, the authors propose a set of methods based on principal component analysis (PCA) to address the problem of recovering precisely the full thermal map from the on-chip thermal sensors when the sensor readings have been corrupted by noise. The authors utilise simulated annealing algorithm to devise method that determines the optimal thermal sensor locations, which can obtain superior results compared with the available literature. On this basis, the authors also propose a practical method for full thermal reconstruction to estimate the accurate temperatures of full chip, which would not need to know a-priori temperature information at each spatial distribution of thermal map. The experimental results confirm that the authors' proposed methods are stable in the case of noisy thermal sensor observations, which can achieve a high fidelity thermal monitoring.