Inter frames in compressed video make transcoding a complicated task since motion information is involved. This is especially true for downscale (spatial resolution) video transcoding in which input motion information cannot be precisely mapped to the output, which causes motion drift. Often carried out in full resolution, the drift compensation loop is computing intensive. In this paper, we model the entire process of downscale transcoding and propose a transcoding architecture that performs all computing-intensive loops in reduced resolution. We show that the computing load is significantly reduced while the quality is preserved.