Milton’s Comment - 3DMark 2000 is used extensively, as far as I can tell. Is this benchmark doing us any favors, now that it is so popular? Or did it simply become obsolete too soon after its release? Is it in danger of becoming the next 3D Winbench 98, where T&L is concerned?
No benchmark program is perfect, and
Mad Onion has admitted imperfection in 3Dmark. No driver is free of hacks
and tricks. No company tells the whole truth. It’s up to the press (on-line
and written) to explain and analyze and to make sure that the average
consumer understands and isn’t tricked, but doing this requires an effort.
Just blindly following what a company whispers in your ear is easy but
it is not correct.
And to end: we should thank Mad Onion
for providing us a benchmark. Where would be without a test program (even
if some of the tests aren’t as ideal as we might want)? Imagine a situation
where the companies provide benchmark programs. Talk about mixed interests.
Do I need to point anyone to the fake benchmark results for AGP throughput
(using a benchmark from our beloved Intel)? It’s all marketing-so some
of its crap-and we need to filter it for the consumer who can’t.
More about HardOCP
Well, HardOCP just
keeps on going. On the 4th of February HardOCP published a document
from Ziff Davis that reads:
Accelerated transform and lighting
-
Tester can chose to use hardware T&L or not
-
All or Nothing: no fall back to software T&L
Now,
HardOCP then starts a rant about this, saying things that seem to indicate
a bit of a misunderstanding on how things work (of course, we honestly
don’t mean anything offensive by this, we appreciate the research they’ve
been doing as much as everyone). Take this:
“What
this shows here, is that when turning off Hardware T&L using 3D Winbench
2000, is that it DOES NOT REVERT to ANY sort of T&L being used, simply
the D3D HAL.â€
Now, simply put, you can’t have 3D
graphics without some form of T&L being used.
Of course the benchmark does T&L, else nothing would show up
on your screen. What happens is simple: Some kind of T&L engine is
used to do the T&L work. The results of this process are then sent
to the D3D HAL (which is the rasterizer – the 3D card that does the actual
drawing of the scene). Now the big question of course is which T&L
is used? Well, this is very obviously the Software T&L of Microsoft’s
Direct3D. While we were unable
to locate the exact quote again, ZD made a statement, via e-mail, saying
that they do not use a custom T&L engine.
So what does
this statement made by Ziff -Davis really mean?
It means that you either use Hardware
T&L or you don’t use it, there is no mixed-mode. This mixed-mode is
the clue. With the Rasterizer you can have some mixed mode which means
things are emulated in the software that the hardware cannot handle. In
the case of 3D Winbench this means that if your card doesn’t support,
say, Tri-Linear filtering then those benchmarks that need Tri-Linear will
be run using the software Rasterizer (which is very slow and results in
a low score for those tests). Now, since 3D Winbench consists of many
tests there will be tests that can be accelerated so scenes that use Bi-Linear
filtering will run fully accelerated. Now, with T&L there is no such
option, either you support T&L or you don’t, there is no option to
support some subpart of T&L. This means that if an imaginary card
would support only Hardware T and not L it would not be able to work using
3D Winbench, 3D Winbench only allows ALL T&L or no T&L.
Now, we do have to question the use of the Direct3D Software T&L engine by Ziff Davis, after all the majority of games comes with custom light and transform engines so a good gaming benchmark should also use one IMHO.
NOTE: Between getting this document proof-read and published there has been an official reply along the same lines from Ziff Davis. You can find it at HardOCP.