Consider the uplink of multi-input multi-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems. When the number of antennas is sufficiently large, the zero-forcing (ZF) detection performed at the Base station (BS) is near optimum to demodulate data symbols transmitted by users over each subcarrier. Nevertheless, it requires matrix inversion to perform the ZF detection especially when the number of users and subcarriers are large. In this work, we adopt singular value decomposition (SVD) based ZF detection and employ power iterative method to reduce computational complexity of SVD. Furthermore, we exploit the fact that the channel matrices of adjacent subcarriers are similar to reduce the required number of iterations in the power iterative method. Specifically, the initial vectors in the power iteration are substituted by the singular vectors obtained for the channel matrix corresponding to the previous subcarrier, rather than the randomly generated vector. It shows through computer simulations that the proposed method reduces the number of iterations by 40%~70%, which significantly reduces the computational complexity in broadband OFDM systems.