Multi-input multi-output (MIMO) is the promising technology for next generation wireless communication system, since it can obtain high throughput and diversity gain on rich scattering environment. However, Maximum likelihood (ML) detection which is optimal bit error rate detector for MIMO system is infeasible due to the complexity of ML when a large number of antennas used together with high order modulation scheme. Therefore, Sphere Decoding (SD), which is the lattice decoding algorithm, is introduced for MIMO system to reduce the complexity of ML detection. SD is an attractive approach to reduce the implementation complexity of ML detector, significantly. But it still has the high complexity. In this paper, two efficient architectures are proposed for SD. One is efficient for small area and high throughput, the other is the architecture for run-time constraint. First architecture efficiently reduces the memory usage and critical path up to the half of the state-of-the-art architecture. To reduce the complexity, an enhance direct Schnorr-Euchner (SE) enumeration scheme which uses the property of quadrature amplitude modulation (QAM) signal constellation is also proposed. The second one finds the ML solution earlier by adopting parallel processing, and it makes the instantaneous throughput more stable with Modified-Schnorr-Euchner(MSE) algorithm.