I've heard the arguments both ways. The tentative conclusion I have made is that the rounding logic is so small that it's nowhere near the highest priority for optimization. My opinion is that we should leave it in and keep things closer to IEEE standard until we have EVIDENCE that it's a good idea to do otherwise. At some point, I want to do a variety of experiments and simulations where we synthesize the design with and without a various combinations of different features, examining the effects on circuit area, static power, dynamic energy, performance, and functional correctness. This way, if we're going to strive to be the "most efficient" GPU (on a Pareto surface in the space of many different efficiency criteria), we can be totally rigorous about our evaluation of well-understood feature trade-offs. The bigger challenge will then be to see if we're clever enough to invent entirely novel features that lead to trade-offs that others hadn't thought of before. That's where I'm afraid of falling down here, and that's one of the many reasons I'm not just doing this all by myself or just among the other researchers at the university: The coolest ideas are going to arise from the interactions of many different people with varied backgrounds.
On Sat, Jun 23, 2012 at 6:31 PM, <[email protected]> wrote: > On Sat, 23 Jun 2012 18:22:05 -0400, Dieter BSD wrote: >> >> Some people are of the opinion that the output of a GPU doesn't >> need to be absolutely correct. "It's just graphics." One can be >> sloppy, rounding results, leaving out things like ECC, and so on. >> >> Sometimes getting the right output matters, even with graphics. >> >> >> http://gizmodo.com/5919149/your-brain-scan-looks-different-on-mac-and-pc >> >> A design that works correctly can be a selling point. Not just >> for medical imaging, but also for those using GPU for design >> calculations, and probably many other uses. > > > Furthermore, testing something inaccurate can be ... too complex. > > Finally, just ask the people who used CRAY and CDC computers :-D > > > > _______________________________________________ > Open-graphics mailing list > [email protected] > http://lists.duskglow.com/mailman/listinfo/open-graphics > List service provided by Duskglow Consulting, LLC (www.duskglow.com) -- Timothy Normand Miller http://www.cse.ohio-state.edu/~millerti Open Graphics Project _______________________________________________ Open-graphics mailing list [email protected] http://lists.duskglow.com/mailman/listinfo/open-graphics List service provided by Duskglow Consulting, LLC (www.duskglow.com)
