This paper presents a novel architecture of a soft MIMO detector. To our knowledge, it is the first soft detector architecture based on the Belief Propagation algorithm. The proposed architecture is composed of a first block that is in charge of the calculation of the Euclidean distances. The second one is dedicated to the exchange of information between the received nodes and the candidate nodes of a larger Joint Factor Graph. Thanks to an appropriate level of parallelism, the associated architecture of the soft detector enables efficient exchanges of information. Therefore, it promotes a good trade-off between error rate performance and computational complexity. It also decreases the latency of iterative MIMO receivers.