Due to the fine delay resolution of typical ultra wideband (UWB) systems, Rake receivers need to combine samples from a large number of multipath components (MPCs) in order to collect sufficient signal energy for reliable decisions. In addition, these samples need to be combined optimally in order to minimize bit error probability. The optimal linear minimum mean square error (MMSE) combining scheme might require inversion of a large matrix depending on channel and system parameters. Therefore, suboptimal algorithms with lower computational complexity but close-to-optimal performance are desirable. In this paper, a low-complexity combining scheme is proposed for that purpose, which divides samples into a number of groups and performs MMSE combining in two steps. Performance of this two-step combining scheme is investigated theoretically and by simulations.