Elmer, I changed my mind again.
developer.intel.com (go to page 6 or 7)
"For FSPEC 2000 the new SSE/SSE2 instructions buy about 5% performance gain compared to an x87-only version"
> I glad you realize the foolish claim that despite having a
> very weak x86 unit the P4 somehow went from only so-so FP
> performance to the highest FP performance ever measured. A
> weak x86 somehow managed to beat Alpha and HP without the
> help of SSE2. Some feat!
Are you still in disagreement with the company that actually published these scores?
Granted, there were performance advancements beyond this. As they suggest in that article, "As the compiler
improves over time the gain from these new instructions will increase."
But at this current moment we do not have any concrete evidence suggesting that there absolutely was more performance eeked out from the SIMD. If there was, and if this has been documented, then I can accept that. Intel's paper there says that the performance improvement from a 1.00GHz Pentium III to a 1.50GHz Pentium 4 is 75%, which means that 66.7% of the improvement is by means other than SIMD, at least at this time.
I did a quick look and tried to find both an early and a recent score for the 1.50GHz Pentium 4 such that as many variables as possible were unchanged. Here are the two examples I found
spec.org
spec.org
spec.org
spec.org
In both cases, the improvement for specfp2000 is between 10 and 12 percent. Other than compiler version, the only change that I could tell was that the newer versions had higher Win2k service pack levels. I do not know if that could improve performance.
Anyway, at this point, the 1.50GHz Pentium 4 is now about 11% faster in specfp2000 than it was previously. If you assumed that the entirety of this improvement is from SIMD, then you would conclude that Pentium 4's implementation of SSE2 adds about 16.5% improvement over Pentium 4's implementation of x87. Of course, there is no guarantee that all, any, some, or none of the improvement (above that initial 5%) is from SIMD, as Intel has not said anything of which I am aware. The likelihood of it being mostly from data transfer related stuff is just as good as the likelihood of it being mostly from SIMD related stuff.
So, from what we can tell, P4.SIMDdp is so far between 5% and 16.5% faster than P4.x87dp. Anywhere in that range is still a very, very small amount compared to the overall 75% boost (83% if you look at the actual scores, though there is at least some hardware change, such as a jump from ATA66 to ATA100 on the IDE level). So suggesting that SIMD is the reason why P4 is so fantastic in specfp is just a way of saying that you haven't really done the research.
-JC
*Edit*
The actual reason why the P4 performs so well in this benchmark suite is another interesting topic, but maybe we'll chat about that later. :) |
