In this paper, we propose an improved sphere decoder (SD) and fixed-complexity sphere decoder (FSD) for multiple-input multiple-output (MIMO) wireless systems in detail. Furthermore, we analyze the performance and complexity of SD, FSD and other popular algorithms (such as Zero Forcing (ZF), Minimum Mean-Square Error (MMSE), Ordered Successive Interference Cancellation based on ZF criterion (ZF- OSIC) and MMSE criterion (MMSE-OSIC) and Maximum Likelihood (ML)) by comparing the Bit Error Rate (BER) and the average detection time consuming. Simulation based on the platform of MATLAB show that: 1) SD achieves the closest ML BER performance for a 4×4 MIMO system using QPSK, 16-QAM, and 64- QAM modulation; 2) the computation complexity of SD and FSD increases with the size of the constellation used; 3) when the SNR is lower than 3dB, the computation complexity of SD is higher than that of FSD; vice versa. FSD achieves the performance of quasi SD while overcoming the two main drawbacks of SD: variable complexity and sequential structure. This makes FSD more feasible for hardware implementation than SD.