The implementation of 3D stereo matching in real time is an important problem for many vision applications and algorithms. The current work, extending previous results by the same authors, presents in detail an architecture which combines the methods of Absolute Differences, Census, and Belief Propagation in an integrated architecture suitable for implementation with Field Programmable Gate Array (FPGA) logic. Emphasis on the present work is placed on the justification of dimensioning the system, as well as detailed design and testing information for a fully placed and routed design to process 87 frames per sec (fps) in $$1920 \times 1200$$ resolution, and a fully implemented design for $$400 \times 320$$ which runs up to 1570 fps.