In particular, a 24b data width takes significantly longer to process than the current tests allow; I suspect all non-power-of-2 cases take longer than their power-of-2 counterparts. Either the implementation needs tightening to give some breathing space or the test needs loosening in these cases.