After post-compression, a smaller rank does not automatically bring faster serving, because issues such as padding, factor access, and generic kernels can mask the retained low-rank structure.