Maximum likelihood (ML) detection is an optimal solution for the multiple-input multiple-output (MIMO) communication system, but the large complexity of ML made it difficult to implement in HW architecture. Sphere detection is an attractive approach to reduce the implementation complexity significantly without compromising performance of ML detector. As a modified sphere decoder, List Sphere Decoder(LSD) is a promising algorithm to guarantee superior performance of BER than Sphere Decoder which uses real-value expanded channel matrix and received vectors is analyzed, designed and implemented using a standard 0.25 um COMS process. Although the decoding latency of this scheme is almost twice the time of the complex-value based LSD, the hardware resource usages of the real-value based LSD are reduced down to root of the complex-based LSD. Also, the pipelining scheme is applied in this design to enhance the maximum clock frequency which is related to the throughput directly. Most of the related implementation is about Sphere Decoder and 4$\times$4 16QAM modulation. In the other hands, this paper implemented a LSD MIMO detector for a system of 3$\times$3 16QAM 3$\times$3 64QAM, 4$\times$4 16QAM and 4$\times$4 64QAM modes. It contains 408K gates and occupied a core area of 4.7mm $\times$ 4.7mm (including pads). The maximum searching latency of this architecture is 160 clock cycles at 4$\times$4 64QAM. Because LSD detects the list of candidates instead of just one candidate for SD, the average detection cycle of LSD is almost 20 times longer than SD. Therefore, the throughput of the designed LSD has to be also 20 times worse than the SD in  with the similar clock frequency. But, because the proposed architecture of LSD uses 3-stage pipelining scheme, the throughput is only about 7 times worse than SD in . Also in sub-section 8.3 showed that when the proposed LSD is implemented with the 90nm CMOS process, the throughput of the designed LSD can be almost satisfied.