> Define "real" applications, Something that produces tangible, scientifically useful results that would not have otherwise been realized without the availability and capability of that machine.
> but to give my guess at your question "But they didn't. Why?" > > One word - cost Well, that's the obvious (and universal) given. But it's not a useful answer in this context. Cost is always a limiting factor. Optimizing capability within the budget envelope is the challenge. Now to be fair, my question was somewhat leading (and my argument is somewhat reduction-to-the-absurd) but what if the system designers has reduced the number of CPU cores per node and used the money saved there to purchase additional GPU nodes. Make the system really CPU light and GPU heavy. You would be left with a something that would potentially have a higher HPL number while maintaining the overall system cost. But why didn't they? Why instead did they spend their money on a things like a custom high-perf interconnect (which tends not to be a limiting factor in HPL performance) and lots of cores on each node? And IIRC 20% of their nodes don't even have GPUs? My point is that while GPUs are certainly a potent tool to use in HPC, trying to draw some sort of universal claim about their efficacy and general usefulness based upon a single contrived benchmark is essentially the same as trying to extrapolate any conclusion from a single data point. Unfortunately many people in the media do not seem to have any reservations about doing exactly that. Take care, and have a great weekend. -b _______________________________________________ Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
