Transforming incoming PostScript into PDF and then back to PostScript is done to do page management (N-up, even/odd pages, reverse order, selected pages, ...) on a PDF data stream (done by pdftopdf), as in PDF you can always reliably tell the pages apart. The incoming PostScript data often comes in non-DSC-conforming form (note that PostScript is a programming language) and so the page management simply does not work, leading to broken printouts or even no printout at all.
Note also that the cost factors are not the expected CPU load, but a way of control to get the preferred filters used if there is more than one solution.
See also http://www.linuxfoundation.org/en/OpenPrinting/PDF_as_Standard_Print_Job_Format Till Martin Pitt wrote:
Hello Till, cpdftocps is a local filter, is this obsolete? Thanks, Martin