Saddle point matrix is the coefficient matrix of the linear system derived from the saddle point problem, which arise from many scientific and engineering applications. The eigenvalue information of the saddle point matrices is usually quite important in practical computation. In this paper, the power method is applied to computing the spectral radius of the saddle point matrices. Furthermore, based on some theoretical results of eigenvalue estimates for the saddle point matrices, a new indicator, which is a function of the extreme eigenvalues of the sub blocks, is proposed to predict the convergence rate of the power method. The numerical results of the experiment on computing the spectral radius of the saddle point matrices derived from the model of the Stokes equation demonstrate that the proposed algorithm and the indicator are both effective.