Yeah quite good
Author: Sharps97 Number: of 40695
Subject: Re: benchmark predictions for 2002 Date: 4/15/02 12:59 PM
Post New • Post Reply • Reply Later • Create Poll Report this Post • Recommend it!
Recommendations: 3
You've been quite busy today Alan, =)
(1) Clock rate scaling: The P4 was introduced in Nov. 2000 at a
frequency of 1.5Ghz, and hit 2.4Ghz in Apr. 2002. The Athlon was at
1.2Ghz in Oct. 2000 and scaled to 1733Mhz in mar. 2002. This means the P4 clock rate has scaled by 60% in the same amount of time it took the Athlon to scale by 45%. The P4 does have the advantage of 0.13u
technology today, while the Athlon is still on 0.18u, however it is
believed that the AMD 0.18u technology makes use of many of the 0.13u
transistor benefits. This is evidenced by the tremendous 250% clock
rate scaling AMD did on the 0.18u process. They are forecasting a much smaller clock rate scaling moving forward onto the 0.13u process.
Interesting. The .13 micron Pentium 4 is using a number of sub .09 micron parts. Intel better be careful to not snack on its future by already using .065 class technology. AMD's .13 SOI better be careful as well, hitting .05 tech in parts of the chips as well. Hmmm...
(2) Benchmark optimization: The improved performance of the Athlon
XP over the thunderbird indicates that the benchmarks are taking
advantage of SSE operations, which the P4 excels at. Code recompiles
have also helped with this effort.
I think it depends on the benchmark. SSE does help some, but Palomino has a few other enhancements that can be related to getting additional performance. I think that some coding has gotten better, and that Athlon XP driver support with video card drivers must be robust. If you are talking games, it may be that XP's raw FPU power and strong integer performance are more than what the P4 can muster.
As far as your little chart goes, I think it is missing some important considerations once you get past 2800+ level of performance (aside from using just Tom's as a basis for comparison). The Clawhammer will have the equivalent of a 333 MHz FSB without the latency associated with the northbridge. This will give a significant boost and makes using a simple linear scaling somewhat questionable. Clawhammer will have a full SSE 2 unit as well. While we've got scaling of model numbers, it seems as though the way they determine what rating is given might be altered by that point (could be less aggressive again).
Anyhow, you probably might want to consider retooling your spreadsheet a bit starting at 3000+ (where I think the CH will intro) up to 3400+. I've provided some speculative specINT and specFP using the same methods as you, basically:
XP_MN sINT sFP Notes
1500+ 577 536 Actual result
1600+ 595 547 Actual result
1700+ 656 604 Actual result, 10% pop from move to KT266A
1800+ 671 615 Actual result
1900+ 701 634 Actual result
2000+ 724 642 Actual result
After this, I figured out the average scaling (excluding the influence of the 10% pop) and scaled forward to see where things would head. To shorten the list here, I skipped a few listings, but the spread sheet still has them.
2100+ 748 655 Estimated
2200+ 772 669 Estimated
2400+ 824 696 Estimated
2600+ 879 725 Estimated
2800+ 938 756 Estimated
3000+ 1163 925 Estimated 64 bit recompile and 333 FSB pop
3200+ 1240 964 Estimated, back to previous scaling percentage
3400+ 1323 1004 Estimated
You can see where performance (in Spec) would get a very large pop just from moving to the CH architecture. Way out guesses for now, but another thing to look back on to see how the actual results turned out. In my model, I didn't account for the impact of SSE 2, but maybe I overestimated the gains for something else. I'm not sure how well this would translate back into raw 32 bit performance, but either way it is something to consider.
|
