Abstract: The power macromodel presented in this paper is built on the word-level statistics of input stream, which can be applied to the power estimation of various Audio/Video DSP chips. In comparison with the fitting macromodel based on bit-level statistics, it achieves better robustness and satisfactorily remains accurate when the signal pattern changes vastly. This macromodel only consumes 1%～0.1% run time of the gate-level simulation and its relative error is no more than 20%. On the other hand, it can evaluate the low power encoding method by distinguishing the different digital presentations.