nVidia release det 43.45.

[BoardBonobo]

They are now available on the nvidia web site. They give good 3DMark03 results and everything is rendered correctly.

Most impressive though is the fact that PS2.0 scores have trebled. So they are up from ~14fps to ~45fps. Are nvidia getting on top of their driver problems?

[FiggyG]

Appears so... but I don't have a GeforceFX to test them on. Now maybe the FX will become competitive :lol:

[Evildeus]

Yeah that guy, from 13 (42.68) to 42 (43.45) and seems to have good quality!

[binmaze]

Post the screen shot of 3DMark03 Game Test 4 'Mother Nature' please.

[binmaze]

A hardware site tested the 43.45 driver and found out 'Mother Nature' IQ still not good.
It seems nvidia giving up FP32...Or 50.xx drivers will do a miracle?

[Evildeus]

A hardware site tested the 43.45 driver and found out 'Mother Nature' IQ still not good.
It seems nvidia giving up FP32...Or 50.xx drivers will do a miracle?
Where?

[binmaze]

I abstained from posting the link...But here it goes...
GT4 Comparison Link (http://www.darkcrow.co.kr/image/tip&tech/4345tech/iq/game4.jpg)

[gkar1]

Clearly lacking overbrights, just as with 42.72. Shame.

[BoardBonobo]

Maybe it was a cloudy day? :D

[binmaze]

Please notify it's from Darkcrow when you post that link on other sites.
I don't have anything to do with that site. Just taking cautions.

[Tim]

A hardware site tested the 43.45 driver and found out 'Mother Nature' IQ still not good.
It seems nvidia giving up FP32...Or 50.xx drivers will do a miracle?

I think it is rather interesting that nVidia chose to release a new non-WHQL driver this raises some interesting questions, such as what is the shader precession?

[Evildeus]

Still :(

http://www.darkcrow.co.kr/image/tip&tech/4345tech/iq/game4.jpg

http://www.darkcrow.co.kr/Review/Review_Content.asp?board_idx=146

[gkar1]

Yep the same ol' 16 bit hackery of 42.72. Pathetic

[Hyp-X]

Most impressive though is the fact that PS2.0 scores have trebled. So they are up from ~14fps to ~45fps. Are nvidia getting on top of their driver problems?

It has likely little to do with driver problems and more to do with the shaders in the test not optimized.

1. While Futuremark supports an alternative PS1.4 path on game test to improve performance on a set of cards, they don't support PS2.X for the PS2 tests. There are a few stuff in PS2.X that can make it much faster.

2. They used HLSL to compile the shader which is fairly far from what I consider good optimizing compiler (sometimes it produces more than 2x the instructions one'd use).

I wouldn't be surprised to find out that hand optimizing these shaders can give the 3x speed up.

[Dave Baumann]

Where did you get the impression that they used HLSL? I asked them this question and their lead developer told me they didn't utilise any HLSL in the production of 3DM03, and they are all hand written assembly.

Also, FYI, Futuremark generally only support a feature in the game tests only if more than one vendor uses it at the time of release. PS1.4 would be supported by both ATI and NVIDIA by the time of 3DM03's release, but only NVIDIA would have supported 2.X

[Hyp-X]

Where did you get the impression that they used HLSL? I asked them this question and their lead developer told me they didn't utilise any HLSL in the production of 3DM03, and they are all hand written assembly.

Found the shader sources in their pakfile.

All the shaders are hand written except the ones used in the PS2.0 test.

Also, FYI, Futuremark generally only support a feature in the game tests only if more than one vendor uses it at the time of release. PS1.4 would be supported by both ATI and NVIDIA by the time of 3DM03's release, but only NVIDIA would have supported 2.X

Like the Nature test in 3dmark2001 :roll:

Not that it matters now as nVidia likely rewrote said shader in 2.X.
(Btw, I'm not talking about a game test...)

[demalion]

Ah, so you're not talking about the nature benchmark figures (which are apparently due to precision demotion).

I think nvidia is optimizing for the synthetic tests to lend validity to the gametest increases, since things were just a bit too obvious before...image quality is harder to discuss than simply pointing to the dedicated pixel shader test and asking why there is such a disparity, and it is the scores that people use for their buying decisions, not the synthetics (unless used as above).

I do think Futuremark should release a patch for 3dmark 03 when the HLSL compiler is updated to support 2.x, but by maintaining the apparent practice of precision demotion for atleast one scoring test in conjunction with specifically addressing the synthetic benchmark people have used to illustrate that something questionable is going on, it looks to me like nvidia's purpose behind this PS 2.0 optimization is to assist in deception.

Hmm...the idea of substituting shader code is pretty questionable, though what they are substituting, if you are correct, does seem to be valid IMO. As I've said before, the problem is that it is apparently a benchmark specific substitution...if they had just updated their LLSL compiler to optimize for their extended features on the fly for all apps, I don't think there would be anything wrong (there seems to be a chance this might be the case, I hope someone will investigate).

I don't know who is to blame for the lack of 2.x support in the compiler (assuming this is still the case), nVidia or Microsoft. To me, it seems consistent with nvidia's Cg initiative and their assumption of market dominance to allow their HLSL control strategy to succeed. The question would then remain which party is promoting the delay in response to that strategy, and I only see benefit to nvidia and the perception of Cg from it, which would also be consistent with efforts to discredit a DX specific benchmark.

All around, it looks to me like the content of the PS 2.0 optimization is likely valid, but their purposes and actions associated with its introduction all seem to be more of the same "dupe the consumer instead of serving them" philosophy they seem to have taken to heart. :-?

[Hyp-X]

Ah, so you're not talking about the nature benchmark figures (which are apparently due to precision demotion).

Yes I'm talking about the PS2.0 feature test.
The Nature scene is clearly lacking something (most probably dynamic range).

I think nvidia is optimizing for the synthetic tests to lend validity to the gametest increases, since things were just a bit too obvious before...image quality is harder to discuss than simply pointing to the dedicated pixel shader test and asking why there is such a disparity, and it is the scores that people use for their buying decisions, not the synthetics (unless used as above).

I think they optimize for the syntetic test to regain confidence of people who didn't want to buy the card because the extremely low score is questioning the "future proofness" of the card.

I do think Futuremark should release a patch for 3dmark 03 when the HLSL compiler is updated to support 2.x, but by maintaining the apparent practice of precision demotion for atleast one scoring test in conjunction with specifically addressing the synthetic benchmark people have used to illustrate that something questionable is going on, it looks to me like nvidia's purpose behind this PS 2.0 optimization is to assist in deception.

Well why should Futuremark release a patch when they said they don't use HLSL at all?

Hmm...the idea of substituting shader code is pretty questionable, though what they are substituting, if you are correct, does seem to be valid IMO.

I don't know if I'm correct I have no ways to check it.
They might just turned it into FX12 "hackery".
Altough appearenly the big advantage of FX12 is that it can work parallely with texture fetches, which are minimal in the PS2.0 test, so I don't see it help them too much.

As I've said before, the problem is that it is apparently a benchmark specific substitution...if they had just updated their HLSL compiler to optimize for their extended features on the fly for all apps, I don't think there would be anything wrong (there seems to be a chance this might be the case, I hope someone will investigate).

There is an inherent problem in HLSL - IHV's can't upgrade it, only ISVs can.
This is one of the fundamental differences with GLslang.

ISV's either include the HLSL code in their program, but the compiler is linked statically (no dll), so a patch have to be released for the program, or they include the compiled code, in which case they have to recompile it and release a patch with the new compiled code.

I don't know who is to blame for the lack of 2.x support in the compiler (assuming this is still the case), nVidia or Microsoft. To me, it seems consistent with nvidia's Cg initiative and their assumption of market dominance to allow their HLSL control strategy to succeed. The question would then remain which party is promoting the delay in response to that strategy, and I only see benefit to nvidia and the perception of Cg from it, which would also be consistent with efforts to discredit a DX specific benchmark.

Well I didn't really care, all that matters is the end result.
If nVidia f*cks up developer relations, they'll end up hand optimizing every shader in every game one-by-one.
If they do it right the developers will support their stuff so they don't have to do it.

But when the game/benchmark runs at the same speed at the end - why should the end user care about it?

(I do care about it, but that's a different story.)

[demalion]

...
I do think Futuremark should release a patch for 3dmark 03 when the HLSL compiler is updated to support 2.x, but by maintaining the apparent practice of precision demotion for atleast one scoring test in conjunction with specifically addressing the synthetic benchmark people have used to illustrate that something questionable is going on, it looks to me like nvidia's purpose behind this PS 2.0 optimization is to assist in deception.

Well why should Futuremark release a patch when they said they don't use HLSL at all?

Well, because you seem to be indicating they are for the PS 2.0 test, of course...I remember the denial as being applicable for the game tests, but I'm not sure they said the same for the PS 2.0 test.


Hmm...the idea of substituting shader code is pretty questionable, though what they are substituting, if you are correct, does seem to be valid IMO.

I don't know if I'm correct I have no ways to check it.
They might just turned it into FX12 "hackery".
Altough appearenly the big advantage of FX12 is that it can work parallely with texture fetches, which are minimal in the PS2.0 test, so I don't see it help them too much.

Well, it uses texture fetches for the turbulence look up which seems to figure prominently in the calculations, so it should be a significant factor, I think. I'd think a detailed turbulence representation might create texture cache misses, and integer processing would seem to be necessary for the nv30 to efficiently use textures as function lookup data in that way...?

As I've said before, the problem is that it is apparently a benchmark specific substitution...if they had just updated their HLSL compiler to optimize for their extended features on the fly for all apps, I don't think there would be anything wrong (there seems to be a chance this might be the case, I hope someone will investigate).

There is an inherent problem in HLSL - IHV's can't upgrade it, only ISVs can.

LLSL wasn't a typo where you replaced it with HLSL above, I meant the run time op code conversion to card instructions by the drivers.

I recognize the advantages of GLslang (though the capitalization protocol seems to give me problems :P ), and I hope nvidia is able to address one cross vendor standard without playing these games. :-?

...
I don't know who is to blame for the lack of 2.x support in the compiler (assuming this is still the case), nVidia or Microsoft. To me, it seems consistent with nvidia's Cg initiative and their assumption of market dominance to allow their HLSL control strategy to succeed. The question would then remain which party is promoting the delay in response to that strategy, and I only see benefit to nvidia and the perception of Cg from it, which would also be consistent with efforts to discredit a DX specific benchmark.

Well I didn't really care, all that matters is the end result.

Well, I'd argue intent matters as well in this case, as it speaks to future decisions and their end results.

If nVidia f*cks up developer relations, they'll end up hand optimizing every shader in every game one-by-one.
If they do it right the developers will support their stuff so they don't have to do it.

Well, the first might impact nvidia card owners, and the second might have a significant impact on non-nvidia card owners. It seems to be demonstrated that time spent by developers doing one thing is time not spent doing something else.

But when the game/benchmark runs at the same speed at the end - why should the end user care about it?

Because it can impact the quality, performance, and repeatability of the end results for all graphics cards.

(I do care about it, but that's a different story.)

Hmm...I'm not sure it is a different story, I just think you're quite a bit more informed about the story than consumers in general...

[Ozymandis]

That's odd... these new drivers "break" UT2k3 16 bit color :?

I wonder if that problem has always existed or what.

[Hyp-X]

LLSL wasn't a typo where you replaced it with HLSL above, I meant the run time op code conversion to card instructions by the drivers.

Oh, I never seen the term LLSL used for the PS Asm -> machine code compiler. I hope it won't catch on.

Yes, actually that's what nVidia changed (that's all they could change).
The question is are they improved the optimization algorithm in general, or did they special cased for this shader...

[Ozymandis]

Wow, they also don't display a proper picture with PowerDVD in hardware-accelerated mode :evil:

Thanks nVidia! :roll:

[mboeller]

Nvidia, The Quack?

Is this real?

Link : http://www.tweakersguide.de/articles.php?show=article&id=30&page=1

Nividia seem to be cheating in Quake III with AA+AF enabled :

driver 41.09 and 2xAA + 2xAF :
http://www.tweakersguide.de/images/articles/4109vs4345_quality/2xAA-2xAF-4109.jpg

the same with driver 43.05
http://www.tweakersguide.de/images/articles/4109vs4345_quality/2xAA-2xAF-4345.jpg


driver 41.09 and 4xAA + 8xAF :
http://www.tweakersguide.de/images/articles/4109vs4345_quality/4xAA-8xAF-4109.jpg

the same with driver 43.05
http://www.tweakersguide.de/images/articles/4109vs4345_quality/4xAA-8xAF-4345.jpg

some will call it a bug, some cheating. Seems Nvidia follows the footsteps of ATi back then in their old times. According to the source this odd behaviour only happens in QuakeIII but no other game. So it could be a bug, but IMHO not cause no other game is used as much for benchmarks as QuakeIII even today.

[Hyp-X]

Nvidia, The Quack?

Is it on GeForce 4 or on GeForce FX ?

[LeStoffer]

Do I detect infidels trying to question The Golden Standard of drivers here?

Blasphemy!

[Tahir2]

It's a bug.
Hang on it can't be a bug since NVIDIA dont make shoddy OGL drivers.
NVIDIA certainly do not cheat in their drivers either.

Hmm...following on logically this leaves QuakeIII as the only suspect, so logically John Carmack has been a naughty boy.

[Unit01]

LOL Lestoffer :D
Thanks for the info Mboeller

[Livecoma]

Does it mention if the driver is in application, quality, or performance mode in that IQ comparison?


That could potentially explain those results your showing...

[Sxotty]

Where can I see some other kind of info on the Quake thing, on an english speaking site, and preferably one that is somewhat trustworthy? Thanks, I am actually quite curious.

[Fuz]

Are these new drivers WHQL certified?

[incurable]

Are these new drivers WHQL certified?
Nope.

[Dave Baumann]

Just installed these on the 5600 Ultra and they appear to make the 8X (not 8xS) and 16X FSAA modes available on this board.

[BoardBonobo]

Dave, does (PowerDVD XP) hardware assisted DVD playback work with the 5600? It is broken with the GF4 in these drivers.

[McElvis]

Just installed these on the 5600 Ultra and they appear to make the 8X (not 8xS) and 16X FSAA modes available on this board.

Can you tell the difference between 8x and 8xS?

Also, what's the performance drop going from 8x to 16x