Optimization method of thermal sensor placement for microprocessor with noise

High-performance processors employ dynamic thermal management (DTM) techniques,which use a set of on-chip thermal sensors to continuously monitor temperatures during the runtime. However, on-chip sensors are inevitably accompanied by a range of noise sources due to fabrication randomness and fluctuations, which greatly affects the reliability of DTM. Therefore, in order to improve the accuracy of thermal monitoring, this paper utilizes principal component analysis (PCA) to approximately reduce dimensionality of the original sample matrix of the thermal image, and combines it with matrix perturbation analysis to propose the optimization method of thermal sensor placement which is based on simulated annealing algorithm. Experimental results indicate that the proposed method significantly outperforms the existing greedy algorithm in the aspects of thermal reconstruction error, Signal to noise ratio (SNR) and false alarm, which can be better applied in dynamic thermal management techniques to achieve accurate thermal monitoring.