|
Futuremark's 3DMark05 - An Introduction |
|
Written by Phyro
|
|
Wednesday, 29 September 2004 |
And also Hexus has posted some
article on the new Futuremark's 3DMark05. If, like many of your enthusiast peers,
you've been staring at the counter on Futuremark's website for the past couple of days, wondering
when 3DMark05 would make its debut, wonder no longer. If you're
reading this, it's past 1pm on Wednesday, 29th of September 2004, and we're
allowed to talk about the biggest deal in PC hardware
benchmarking.
 Futuremark's 3DMark05 - An Introduction
Futuremark's use of DST
acceleration, in combination with single-cycle PCF filtering that's used for
sampling of the DST, a method that has no part in PC-based DirectX and lies far
outside Futuremark's
previous claims that they'd not use vendor-specific extensions to render their
3D benchmarks on the PC, is the sticking point for me. That it's enabled by
default on hardware that supports it is even more disagreeable. Combine that
with the fact it reduces image quality (in my eyes) in the quest for
performance, and we're back to a point we had during 03's lifetime, but this
time it's supported by Futuremark and so it's valid in some respects.
When you
also understand that Futuremark declined to put 3Dc support into 3DMark05,
something that'd help reduce vertex load on a vertex-fetch-limited benchmark,
for the precise reason that it's not a part of DirectX, is galling. Neither is
nVidia's DST or PCF method of
PSM acceleration, so why's that in there?
3Dc is a ratified part of WGF
1.0. It'll at least be in DirectX at some point in the future. The same can't be
said of the nVidia acceleration of PSMs in 3DMark05.
That
it's a decent GPU-bound test that'd slot right into some kind of performance
evaluation as far as I'm concerned, when partnered with valuable game-based
benchmarks and other synthetic tests is invalidated by the current state of the
benchmark and the technologies it includes (and discards) and how they're
presented as sensible defaults.
DSTs need to be off when comparing nVidia hardware to ATI hardware using 3DMark05,
something I'll be doing next week.
The various sub tests are great. The
fillrate tests for example are better than they were in 3DMark03,
and the batch size tests do indeed highlight issues uploading small vertex
batches to the hardware, an issue in current drivers on modern
hardware.
But the issues with the vendor-specific extensions and their
effect on the benchmark make it hard to recommend. We have a handle on its
performance characteristics, essential in explaining how it uses a GPU for its
performance, but that's not enough maybe.
It's a very pretty way to show
off modern hardware, and Macci et al will love to race to 10K and beyond, but
its use as a serious benchmark is let down by use of an accelerated method of
PSM rendering that's yet to show its face on anything other than XBox and which
isn't a standard part of DirectX, leaving its means as a rendering technique on
PC games in the future in doubt, calling into question the reasons why that
method of PSM rendering was ever included in the first place. Performance
certainly seems to be one factor.
It remains to be seen how it'll pan out
in the future, after all they're not called Futuremark for nothing, but the reasoning behind the way it
was designed still needs to be investigated...
|
NiBiTor v6.06
