In this paper, we propose a new idea based on sphere decoder (SD) to realize the tradeoff between the performance and the complexity of detection algorithm for MIMO systems, by decomposing a long signal vector required to be detected into several correlative short segments in the process of detection. It is visually named as split-tree SD since the set of all possible vectors can be represented by a tree structure. The simulation results of splitting at the different layer of the tree show that it is flexible and cost-efficient to trade the BER performance for the greatly reduced computation complexity. Finally, through a comprehensive comparison in a figure of BER vs. complexity, we obtain an adaptive split-tree SD algorithm, which has a better performance with a lower complexity than the traditional V-BLAST detection at the same time