Intel Corp (INTC)

[wbmw]

Re: Everything fell apart in late 2004.Why ? Because of lack of progress with IPF's top frequency.

>> BS! It is really quite simple, GFLOPs per dollar. Cheap
clusters of 2s nodes have taken the HPC world by storm
and the choice of processor is determined by GFLOP/
dollar with TDP an important second criterion.

But isn't this similar to what Savantu is saying? Intel and HP have been steadily decreasing the costs of Itanium through economies of scale and volumes, but over the last few years, GFLOPs has stayed relatively constant, doubling once with the introduction of Montecito. Meanwhile, x86 cores have seen dramatic increases in GFLOPs.

Look at 2004 to 2008:

For x86, Intel has gone from single core Noconas at 3.6GHz with 2 DP FLOPs per clock (7.2 GFLOPs) to quad core Harpertowns at 3.2GHz with 4 DP FLOPs per clock (51.2 GFLOPs). That's a 7.1x improvement.

For IPF, Intel has gone from single core Madison at 1.6GHz with 4 DP FLOPs per clock (6.4 GFLOPs) to dual core Montvale at 1.66GHz with 4 DP FLOPs per clock (13.3 GFLOPs). That's a 2.1x improvement.

Yes, it's true that IPF gets more real world performance per FLOP, but it's not hard for x86 to look more attractive from the standpoint of more cores, operating at higher frequency, without exceeding the same thermal limits. And in fact, Harpertown can run at 80W if you lower the frequency to 3.0GHz, while Montvale is what... 105W @ 1.66GHz? And x86 is certainly less expensive per system.

Fast forward to Tukwila, and it's obvious that Intel will improve IPF's value prop considerably. Not only will they double the cores and increase the clock, but they'll add plenty of memory bandwidth to help bandwidth constrained applications, of which there are many. That's fine, but won't Intel still be ahead with Nehalem? That's an interesting point, from which I have never been able to form a definitive opinion. Intel's x86 efforts seem to be on a fast trajectory, and I wonder if IPF will ever vault ahead.