herb, now it appears that ddram is not living up to the hype
Well, lets see which platform is performing the best, P4 with PC800 RDRAM or Athlon TBird with DDR
tomshardware.com
Sysmark 2001 (a benchmark which runs a highly P4-optimized application, Windows Media Encoder in the background during most tests)
Athlon 1.33 with DDR is 16.9% faster than P4 1.7 with RDRAM
A 13.3% increase in P4 clock speed (1.5 to 1.7) increases performance 7.8%. Slope is 0.59. At that rate, a P4 2 GHz will still be slower than a 1.2 GHz Athlon. I guess that Rambus memory doesn't scale so well, huh?
Is it the memory, or is it the CPU? Hard to tell from Tom's benchmarks, but the benchmarks at zdnet.de and surrounding pages are ideal for evaluating the scalability of the platforms. ZD.de did benchmarks using 266 DDR between 1000 and 1533 MHz for the Athlon, and they did P4 benchmarks using PC800 between 1000 and 1700 MHz. I'll ignore the 1000 MHz P4 benchmark results and calculate the slope of the benchmark result from 1300 to 1700 MHz on the P4, and between 1000 and 1333 MHz on the Athlon. Ideally, the Athlon result would increase by 0.3333 and the P4 result by 0.3077 going from 1000 MHz to 1333 MHz and from 1300 MHz to 1700 MHz, respectively. The ratio of the actual benchmark increase to the ideal benchmark increase is refered to as the platform efficiency in the table below and is a kind of "figure of merit" for the platform.
In this particular benchmark comparison, the Athlon 1.33 won 14 and the P4 1.7 won 9.
The Athlon "wins" are shown in bold.
...........................................................Efficiency Rating = %performance increase / %clock increase
Benchmark..........................Athlon....P4....notes
ZD Bus.Winstone 99..0.261......0.300....But even 1.0 GHz Athlon is faster than P4 1.7
ZD Bus.Winstone 01..0.584......0.623....ditto, less extreme Athlon 1.0 ~= P4 1.6
ZD Cont.Creat2000..0.445......0.588....ditto, less extreme Athlon 1.0 ~= P4 1.3
ZD Cont.Creat2001..0.509......0.677....ditto, less extreme Athlon 1.0 ~= P4 1.6
ZD HiEndWinston99..0.462......0.676....Athlon 1.0 ~= P4 1.6
3D Studio Max 1.1......1.109......1.044....! Athlon 1.0 ~= P4 1.6
3D Studio Max 4.0.....0.963......1.083....Athlon 1.0 ~= P4 1.5
ZD i-Bench HTML.......0.755......0.448....Athlon 1.0 ~= P4 2.0
ZD i-Bench PDF.........0.652......0.612....Athlon 1.0 ~= P4 1.9
ZD i-Bench Java.........0.978......1.757....Obviously P4 1.7 is more optimized for this than 1.3
ZD i-Bench JavaScr..0.905......1.233....Again unbelievable scaling but Athlon 1.0 ~= P4 1.8
ZD i-Bench Flash......0.597......1.032....A1~=P1.4 Efficiency Rating using 1.3 and 1.5 is only 0.581
ZD i-Bench Shock......0.122......0.106....Athlon 1.0 ~= P4 1.1; benchmark not CPU dependent
MadOnion3DMark01..0.230......0.366....Athlon 1.0 ~= P4 1.4
Vulpine GLMark1.1p..0.081......0.153....Athlon 1.0 ~= P4 2.7; benchmark not CPU dependent
Quake3 ArenaNV15....0.497......0.623....Athlon 1.33 ~= P4 1.6 Quake loves Intel!
AquanoxAquamark...0.438......0.604....Athlon 1.0 = P4 1.5, Athlon 1.33 ~= P4 1.9
SpecViewAwadvs04..0.180......0.012....All P4's 6% faster than Athlon 1.33
SpecViewDRV-07......0.234......0.167....Athlon 1.33 ~= 3% slower than P4 1.7.
SpecViewDX-06.........0.216......0.490....P4 1.7 40% faster than A 1.33 + scales better
SpecViewLight-04.....0.390......0.614...P4 1.7 20% faster than A 1.33 + scales better
SpecViewMedMCAD.0.354......0.328....A1~=P1.8 Athlon 1.33 18% faster than P4 1.7 + scales better
SpecViewProCDRS...0.363......0.525....P4 1.7 16% faster than A 1.33 + scales better
In the 14 benchmarks for which Athlon 1.33 was faster than P41.7, P4 was scaling better than Athlon in 10 of them.
Average Athlon platform efficiency rating = 57%
Average P4 platform efficiency rating = 64%
However, even though it appears that the P4 platform is more scalable (efficient), the closeness of the numbers really indicates that the Athlon platform is MORE scalable on these benchmarks. In 10 of the benchmarks, the Athlon 1.0 GHz DDR was already as fast as 1.6 GHz P4 or better.
Note that in a few of the Internet benchmarks, the benchmark performance for the 1.7G P4 seemed anomalously much better than for the 1.5, i.e., the performance actually increased more than the clock speed, so in computing averages, any number greater than 1 was truncated to 1.
In the 9 benchmarks won by the P4, the P4 was scaling better than the Athlon in 6 of them. Here's the averages:
Average Athlon platform efficiency rating = 36%
Average P4 platform efficiency rating = 43%
Using my prior argument, I should say that the P4 platform is astoundingly efficient since it manages to be more scaleable vs. MHz even for benchmarks that it "wins." But what I find interesting is that the P4 seems to only win benchmarks that are relatively CPU-independent, while Athlon wins benchmarks that are highly scalable with frequency.
That has some other implications -- mainly that the Athlon 4's seemingly small 10 to 15% performance improvement will go a long way to closing the gap on the benchmarks Athlon loses, if, as I suspect, it increases performance and scalability in memory bandwidth-dependent, and not CPU speed dependent, benchmarks. A 10% improvement in the Athlon's "weak" benchmarks is equivalent to a CPU speedup of 25%.
For the 50% of benchmarks in which a 1.2 GHz Athlon is already ahead of a 1.7 GHz P4, the P4-like improvement of pre-fetch and SSE will probably make little difference -- because these features didn't help the P4 any.
PS - Someone could go through these benchmark results and extrapolate 2 GHz P4 performance using the 1.7 and 1.3 numbers and compare that to the 1.533 GHz Athlon numbers (not bus-overclocked) included in the results. Would the results be better or worse than 14/23?
Petz |
