UC Riverside

This paper proposes a new thermal modeling method for package design of high-performance microprocessors. The new approach builds the thermal behavioral models from the given accurate temperature and power information by means of the subspace method. The subspace method, however, may suffer predictability problem when the practical power is given as a number of power maps where power inputs are spatially correlated. We show that the input power signal needs to meet some dependency requirements to ensure model predictability. We develop a new algorithm, which generates independent power maps to meet the spatial rank requirement and can also automatically select the order of the resulting thermal models for the given error bounds. Experimental results validates the proposed method on a practical microprocessor package constructed via COMSOL software under practical power signal inputs.