<Hammer is not going to get 25% performance in every single application, and Pentium 4 is not going to scale 100% in every application.>
I never claimed it would. My analysis is valid as long as it averages 20 to 25% performance improvement vs. Palomino ro TBred (at same clock speed). You seem to think that 20-25% is some kind of "best case" or "almost best case" improvement. The Quake III Arena benchmark for Clawhammer proves you wrong.
tecchannel.de
Here the improvement of Clawhammer vs. Athlon is 41%. Will it be that good, on Quake III Arena when comparing at 2 GHz instead of 0.8 GHz? Probably. In general you should expect a newer architecture to scale better than an older architecture. Which means that the Quake III Arena result for a 1667 MHz Clawhammer should be better than 210*183/130, better than 296 FPS. That's the performance level a P4 2.53/533 gets with a faster GeForce4 card. 2 GHz is 20% higher, so a 2 GHz Clawhammer is, IMO, unlikely to perform much better than a 3.06 GHz P4. This is quite consistent with my original post wbere Quake III was one of the benchmarks on which a 2 GHz CLawhammer would NOT reach the 3.4 GHz P4-performance level.
But I digress. My main point was that AMD's "20-25% greater IPC than TBred" estimate should be taken as an average across diverse application benchmarks, not a maximum.
You try to determine a single "Pentium 4 scalability constant" by drawing a trend line given only 2 data points, and then averaging that across a number of applications whose bottlenecks aren't even in the CPU (3DMark, for example).
Have you ever heard of the design of experiments according to the Taguchi Method? Yes, there are a lot of variables that determine the runtime of a benchmark and we are varying all of them in a non-orthogonal manner and trying to come to conclusions about the results. The amazing thing is that you can come to conclusions in such situations.
If you want to do the "experiment" a different way, do this:
First, estimate what each benchmark result would be for a 2 GHz TBred. Say, take 1.8*(Performance at 1.8 - Performance at 1.6) + Performance at 1.8.
Second, add 20% or 25% to these numbers.
Third, compare the result to the top overclocked P4 (a 2.53 GHz P4/533 overclocked to 2.8 GHz). This overclocked P4 is equivalent to a 3.06 GHz P4 with PC1066 RDRAM, because the 2.8 GHz P4 has overclocked everything - memory, hard disks, etc.
And for another thing, if an application is more affected by the video card than the CPU, then you can't use it with your averaged scalability constant, and then assume that Hammer will still get the full 25%
Yes I can, because there will be just as many applications for which 20% or 25% is to small a number to use -- for example, applications where memory latency is the key limiting factor.
Petz |
