os that rather uses the gpu?
jd1008 at gmail.com
Sat Jul 17 02:35:48 UTC 2010
On 07/16/2010 04:49 PM, Robert Myers wrote:
> On Fri, Jul 16, 2010 at 6:36 PM, Christofer C. Bell
> <christofer.c.bell at gmail.com <mailto:christofer.c.bell at gmail.com>> wrote:
> On Thu, Jul 15, 2010 at 10:56 PM, Robert Myers
> <rbmyersusa at gmail.com <mailto:rbmyersusa at gmail.com>> wrote:
> This is a technical forum, not a debating society. Science is
> awash already in useless flops.
> You're showing your ignorance, Robert. FLOPS is a very useful
> measure and more is always good, but I don't think you know what
> it means. It means FLoating Point Operations Per Second. It's
> nothing more than a measure of computational speed and says
> nothing about the accuracy of the models being computed.
> In one of these long-running public debates, I pointed out to
> a national decision-maker how relatively meaningless and
> frequently wrong NOAA's hurricane season predictions have
> been. Guess what? The most recent forecast was wrapped in
> all kinds of weasel words. That won't stop them from drawing
> their paychecks and burning megawatts producing useless pictures.
> If the forecast is wrong, the either the model was inaccurate or
> the input data was wrong. It says absolutely nothing about the
> value of computing the answer fast (what FLOPS provides). While I
> totally get you're going to somehow try to refute the idea that
> "having a fast computer is good" and thus, I can only imagine,
> make the case that "a slow computer is even better," you're not
> going to convince anyone.
> Whether you like it or not, and whether, even more important, you
> understand it or not, the architecture of a computer has a big impact
> on the kinds of models you can implement with any degree of efficiency
> on the computer. Hanging lots of flops on inadequate bandwidth is
> popular because you can get a high ranking on the Top 500 list with a
> minimal expenditure, but it also means that you can't do an important
> calculation like a global FFT at anything like reasonable efficiency.
> Flops are almost free. Bandwidth is expensive. Unfortunately, what
> is almost free (flops) can be nearly useless for, say, an FFT unless
> you have the bandwidth to move data around globally fast enough.
> Skimping on global bandwidth so that you can't do global operations
> effectively forces you to approximations that require only local
> computation, that's the kind of modeling we get, and the fact that
> there is an inherent disconnect between the way the equations behave
> and the way the discretized model behaves gets swept under the rug.
> Nonlinear systems interact globally at all scales at every time step.
> If your discretization scheme forces an inaccurate representation of
> that global interaction, then you shouldn't ever trust what's coming
> out of your simulation of nonlinear systems.
> Attempting to simulate hurricanes is about as nonlinear as it gets.
> When your weatherman on TV knowingly talks about tropical waves as a
> cause of concern, he's talking about a large scale flow pattern that
> might or might not go through some tortuous nonlinear processes that
> will turn it into a hurricane. Modelling the interaction of hugely
> different physical scales accurately is essential to accurate
> prediction, and modern "supercomputers" are notably and chronically
> short of the bandwidth required to do so.
> Anyway, technical forum or not, it's a forum made up of people and
> that makes it a community. To get along in a community, it helps
> to not be an asshole. While you may not care about being an
> asshole, do note that "getting along in a community" does
> influence if anyone takes you seriously. And you seem very much
> to want to be taken seriously.
> Calling other people names helps to build a community? From your
> pompous and condescending tone unbacked by actual knowledge, I'd be
> inclined to label you in the same way you have chosen to label me,
> with the addition of the word "ignorant" in front of it, but I don't
> think name-calling and labeling ever helps.
So, what would you say is/are the class/classes of problems that would
benefit greatly from a high flops gpu, but without the sort of bus
bandwidth you would like to see?
More information about the users