I see... I suppose the efficiency of the presumed software-based floating point library can be factored along with the parallelization then.