Multiple-input multiple-output (MIMO) significantly increases the transmission speed of a communications system by employing multiple antennas on the transmitter and receiver. To get maximum performance from a MIMO system, you need a computationally intensive search-based detector. To address this challenge, custom designs on ASIC or FPGA are usually used in the receiver. This challenge is compounded by the receiver's need to handle different configurations for typical industry standards. The graphics processor unit (GPU) is a low-cost parallel programmable coprocessor that provides extremely high computing throughput for many signal processing applications. However, careful architecture-aware software design is required to take advantage of the performance offered by the GPU. We propose a new soft MIMO detection algorithm, multi-pass trellis traversal (MTT), and demonstrate that we can achieve ASIC/FPGA-like performance and handle different configurations in software on GPUs. In many wireless systems, a channel decoder such as LDPC is combined with a soft-output MIMO detector at the receiver to maximize performance...