Futuremark: 3DMark06

[Hubert]

Well, there must be a reson for X1800 XT being slightly slower than it's direct opponent. (7800 GTX 256 Mb) I can't think of anything else (being an outsider) that simply put, the 3DMark 2006 is taylored more around Nvidia featureset. Of course FutureMark says it's around the DirectX API, but important is, that the theoretic discussion about which SM 3.0 approach is better has come down to test results' level.

But than again, on a second thought , it might have to do with no proper driver optimization, also. After all, the difference is not that great.

 

[Lux_]

why EB says one of the IHVs isn't happy with the way the software has been written?

I think it's because of this:

ATI has to read and write to a 32-bit depth buffer, and also write to a 32-bit color buffer for a total of up to 96-bits per pixel of bandwidth.

And it has to be done on rendering every shadow (correct?)

Edit

FutureMark says it's around the DirectX API

Extremetech's conclusion is quite informing in this regard. Yes, "tailored around" it seems to be. Community outreach, maybe ?

 

[Jawed]

There is branching in the pixelshader in 06 (I can see branch instructions in the shader dump), but (and I'm guessing here without poking it too much) not to improve performance to any reasonable degree.

So static branching then, SC:CT style

Jawed

 

[Joe DeFuria]

So static branching then, SC:CT style

Jawed

If that's the case, then I'm very disappointed.

This thread as a point of reference:

http://www.beyond3d.com/forum/showthread.php?t=27446

So that just leaves dynamic branching as allowing significantly new types of shaders [relative to PS 2.0]. The rest of the improvements, while good, pretty much only allow minor programming ease and execution efficiency improvements.

I would have really hoped that, even if it's not a test that contributes to the overall score, that Futuremark would have worked in some test for dynamic flow control in the pixel shaders. Both major IHVs support it for cryin' out loud...

 

[Dave Baumann]

I think it's because of this:

And it has to be done on rendering every shadow (correct?)

ATI are suggesting that this is 6 passes per frame as well - Nick, is that the case?

 

[Joe DeFuria]

Extremetech's conclusion is quite informing in this regard. Yes, "tailored around" it seems to be. Community outreach, maybe ?

This is one of the more troubling things:

....A GeForce 6200 supports SM3.0 but not floating point blending. We still consider that a SM3.0 card, and yet it uses the formula for SM2.0 cards because it does not support the FP16 blending required for the SM3.0 + HDR tests. Why the discrepancy? Why doesn't the GeForce 6200 just produce no score? The rules are different when AA is enabled than they are when it isn't.

If there's one thing that irks me, is inconsisency with the application of "rules."

 

[Razor1]

Well, there must be a reson for X1800 XT being slightly slower than it's direct opponent. (7800 GTX 256 Mb) I can't think of anything else (being an outsider) that simply put, the 3DMark 2006 is taylored more around Nvidia featureset. Of course FutureMark says it's around the DirectX API, but important is, that the theoretic discussion about which SM 3.0 approach is better has come down to test results' level.

But than again, on a second thought , it might have to do with no proper driver optimization, also. After all, the difference is not that great.

Not really still dling 3dmark06 but if its now pixel shader limited unlike before, even the 7800 gtx 256 should be faster then the x1800xt.

 

[Geo]

Here's a direct link I had good luck with for those of you still struggling: http://www.webnallen.se/download/3DMark06_v102_installer.exe Took a few seconds to start, then good speed.

 

[Neeyik]

According to the documentation, either PCF or FETCH4 can be used. I believe the 'Disable Hardware Shadow Mapping' feature turns off both of these if selected.

FETCH4 is only used if the hardware supports DF24 though.

 

[Hubert]

to Razor1

So you say it's not pixel shader limited. Than what ? Vertex shader limited ? I am eager to read Mr. Baumann's review.

Note, the performance gap between X1800 XT and 7800 GTX 256 widens in the final graphics test, namley Deep Freeze (if I recall it's name right)

 

[Razor1]

Oh, so you say it's not pixel shader limited. Than what ? Vertex shader limited ? Iam eager to read Mr. Baumann's review.

2 tests of 2005 was cpu limited so it skewed the scores a bit

 

[Jawed]

I wonder if the final test is primarily HDR FP-blending limited and therefore not vertex nor pixel shader limited.

Jawed

 

[Razor1]

Here's a direct link I had good luck with for those of you still struggling: http://www.webnallen.se/download/3DMark06_v102_installer.exe Took a few seconds to start, then good speed.

Ah thx Geo

 

[Spaceman-Spiff]

Anandtech's review: http://www.anandtech.com/video/showdoc.aspx?i=2675&p=2

nVidia seems to be leading in the benchmarks.

 

[digitalwanderer]

Extremetech's conclusion is quite informing in this regard. Yes, "tailored around" it seems to be. Community outreach, maybe ?

Ok, this bit just struck me as wrong:

Nvidia likely believes that having no score is better than having a low score. According to Tony Tamasi, Nvidia believes that an "NA" (not applicable) is the correct way to report the score.

 

[Mariner]

Quoting geo from 13th January:

You entirely left out the "FM are cheating for [insert IHV here]!" threads.

Oooh, I can't wait for these to start. A 3DMark launch just wouldn't be right without some controversy!

 

[Hubert]

"Nvidia likely believes that having no score is better than having a low score. According to Tony Tamasi, Nvidia believes that an "NA" (not applicable) is the correct way to report the score. "

Who ? Rollo Tomasi ? Bad joke, I know. Sorry.
Well, if NA, than wouldn't that mean that Nvidia GPU's are not SM 3.0 capable ? After all, they fail some SM 3.0 tests. Wait, someone mentioned that AA is not part of DirectX ... I am confused.

Anyway, the existance of a vertex texture fetch scenario, which will not run on Ati hardware seems to have the sole reason to show this handicap to the masses. Why not include it into the graphics tests if so important ? What other relevance can such a test have?

 

[Nick[FM]]

Hey guys!

Sorry for being rude and only posting a short post here (and copy & paste from our Whitepaper, which is available here: http://www.futuremark.com/companyinfo/3DMark06_Whitepaper_v1_0_2.pdf) but I am swamped with emails and work!

Since someone asked if we use DFC in 3DMark06, here's what SM3.0 features we use in 3DMark06:

Key features of ShaderModel 3.0 used in 3DMark06:

- vPos Register
- Derivative Instructions
- Dynamic Flow Control
- Large number of interpolators
- Large number of constants
- Large number of instruction slots
- Texture instructions with explicit LODVertex
- Vertex Texture Fetch (Required in the Shader Particles feature test)

3DMark06 uses all key SM3.0 features, except for vFace Register.

The N/A score for non-supported AA (or AA quality levels) is from our point of view better, than either letting HW without support for AA run all tests without AA and get a score based on that (non-comparable) or run 2/4 tests with AA (non-comparable) and get a score based on that. The use of the graphics sub scores are as useful as the 3DMark score when comparing ex. AA performance. In fact, if you want to know pure GPU power, the sub scores are very useful, since the 3DMark score now takes in count the CPU scores as well.

Again, sorry for such short & quick post! I'll be back asap!

 

[Apple740]

Nick, is FP blending/filtering being used for HDR, or only FP blending? In case of blending/filtering, how is the impact of filtering with the R520 (because that card doesn't support FP filtering in hardware).

 

[Hubert]

That's a good argument. There's no way of doing it to have a meaningful score.