FirePRO S10000 currently
goes in at No 2 of the Green 500 via the University of Frankfurt.
Coincidentally, that is the same system as mentioned already in in the GK110 thread.
But what's wrong with AMDs FirePro? There's only one system in the Top100 and it's efficiency is an abysmal 23 percent with no power figure given.
http://www.top500.org/system/177996
But they corrected the performance up (421.2 TFLOP/s now, peak stays at 1098 TFLOP/s) and included the power usage (a modest 179 kW).
As assumed over there, the number of FirePro cards is quite low (420). Who knows how they counted the cores (38400)? Is the given peak (~1,1 PFLOPs) even correct?
The 420 FirePro cards have a peak of 620.9 TFLOP/s. As there is probably 1 card per node and the Xeon E5 supports 2 CPUs (16 cores) per node this would give us 6720 CPU cores @ 2 GHz with 8 FLOPs per cycle and core equaling only 107.5 TFLOP/s for the CPUs and 728.4 TFLOP/s for the whole system (and an efficiency of ~58%, what is believable). The core count would also be somehow close if one counts 56 cores (CUs) per FirePro and adds it to the CPU cores. But 30240 cores are still too short of the claimed 38400 to explain the difference with a few IO nodes. And the power consumption of 179.8 kW for 420 nodes each with a FirePro S10000 in it sounds believable (428W for each node, quite okay for a node containing a 375W TDP card plus dual 8 core CPUs plus some overhead for the network infrastructure; I wonder what would have happened with fully enabled Tahitis at 725 MHz and a slightly further reduced voltage
), but not if one would add so many CPUs to account for the balance to the 1.1 TFLOP/s claimed peak.
Anyway, with the recent update of the
green500 list Titan isn't in the #1 spot as claimed (it's now 3rd). Second is this Xeon+FirePro cluster and the first spot takes a small intel only system with Knights Corner cards.
Edit:
AMD already advertises S10000 as one solution to many problems - including that you basically can kill two birds with one stone.
Yes. But comparing it to K10 or K20 individually, it doesn't look that good anymore. So it's not compelling to someone, which has a clear preference for DP or SP performance and doesn't value the performance with the other precision too much. Imagine a 25% higher performance per Watt (at least from the spec side, it could be that the 375W of the S10000 are never really used and it just happens to be slightly over 300W in any realistic scenario) and how this would change the comparison. I said it before, one would sit just 3% behind in SP performance/Watt (with higher absolute performance) against K10 while being 25% in front comparing it to K20. With DP performance it's no competition against K10 and the performance/W would be just 5% behind a K20 (while providing higher absolute performance and bandwidth [just counting the upsides here]). So for people interested in SP performance it's a wash against K10, for DP against K20 and for people interested in some mix one would have had an actual advantage.
Either way, a lower TDP couldn't have been a bad thing.
