Sparse code multiple access (SCMA) is an efficient non-orthogonal multiple access (NOMA) technique designed to address the demands of wireless communication, particularly in systems with massive connections. In this paper, a novel belief-selective message passing (BsMP) algorithm is proposed for the SCMA detection. By circumventing the lattice point with relatively small reliability, the BsMP achieves a lower computational complexity compared with the state-of-the-art (SOA) detector. Numerical results show that for a 6-user regular SCMA system, the proposed BsMP achieves the comparable error performance with the SOA detector while saving about 30% computational complexity. Besides, the hardware architecture of BsMP is also developed, demonstrating the feasibility of the hardware implemention for the proposed BsMP detector.