New Dawn ultra high quality patch

[Uttar]

Hey everyone,

I've just finished my Dawn Ultra High Quality patch.
The patch simply replaces all shader files with full FP32 equivalents ( the original files are a mix of FX12 and FP16 )

This is 100% compatible with the nude patch, however the installation procedure for it when using the nude patch is very specific. Here's a part of the readme:

Usage:

- Extract contents of this ZIP-file to [...]\Nvidia Demos\Dawn
- Run HQ-on to increase the IQ. Please note that the nude patch should NOT be enabled when doing this ( use nude-off is required )
- If you got the nude patch installed and want her nude: Run nude-on.
- Run nude-off ( if you got the nude patch installed ) and then HQ-off to return Dawn to original IQ state. You may use nude-on again later if you want to.

A word of warning first: this patch KILLS performance. Even a GeForce FX 5900 Ultra can't run it at 25FPS+!

The following numbes are on a GeForce FX 5900 Ultra, thanks MikeC for the testing!
1024x768:

Default Shaders - 29fps

FP16 - 27fps

FP32 - 18fps



1600x1200:

Default Shaders - 27fps

FP16 - 25fps

FP32 - 17fps

The FP16 version is not publicly available, please feel free to contact me to get it if you want to test it.
The FP16 numbers there also proof that the NV35 got no legacy FX12 hardware: it ues FP16 when asked for FX12 ( although it might still have some very limited FX12 functionality explaining the slight performance hit of FP16 )

Anyway, here's the link to the patch:
www.notforidiots.com/DawnHQPatch.zip

Any feedback?
The IQ difference should be VERY slight, maybe not even visible by the bare eye. Some comparaison screenshots would be nice. Can't do that with NV30 Emulate...


Uttar

[Hanners]

No point ATi users with the OpenGL wrapper trying this I guess? ;)

[Uttar]

Originally posted by Hanners
No point ATi users with the OpenGL wrapper trying this I guess? ;)

Yep, completely futile. ATI hardware supports FP24, not FX12/FP16/FP32.


Uttar

[Bopple]

I'm just a layman. Out of curiosity, I got a question.
Does FX12 data converted into FP32 result in improved IQ?

[Uttar]

Originally posted by Bopple
I'm just a layman. Out of curiosity, I got a question.
Does FX12 data converted into FP32 result in improved IQ?

Err, don't exactly understand what you mean there...
The original demo was in FX12/FP16, and now it's fully in FP32. It's not "data converted into FP32" - it's simply a different level of calculation precision.

Uttar

[Bopple]

Maybe I have misunderstood precisions.
I've thought FP32>FP16>FX12 on data-wise. Therefore, if FX12 data types containing less information than FP32, what's the point of calculating it in FP32 mode?

Or...the original color data have nothing to do with precision calculations?

[Dazz]

If the performance hit is so big why did nVidia make their GPU's support it?

[Grrrpoop]

Originally posted by Dazz
If the performance hit is so big why did nVidia make their GPU's support it?
So that they can say "We have the FIRST consumer gfx card with 128bit precision Cinematic Rendering rah rah!"

It's like the support for extremely long shader instructions, and you would assume there's a lot in the Dawn demo but no..

If we ever saw a demo which utilised everything the FX "supports" we'd think we were looking at a static image ;)
To balance that statement out tho, the same can be said for the R350, with the exception that it actually performs very well at FP24 all the time..

And despite Dawn not running in FP32 mode as standard, it still looks superb, which to me says FP32 is just bragging rights and not currently essential. I'd like to see some screenies of FP16 Dawn vs FP32 Dawn :)

I think ATI hit the sweet spot with FP24 tho.

[Uttar]

Originally posted by Dazz
If the performance hit is so big why did nVidia make their GPU's support it?

Because if you used it to make a small, essantial, part of the scene look good, then it isn't too slow.
Heck, that's what they're doing by default in Dawn: the eyes, for example, are done in FP32.

I must admit I'd like some comparaison screenshots between default and FP32 too, but NV30Emulate does not correctly indicate real IQ I think ( CPUs don't even support FP16... Only FP32 and FP64 )

Uttar

[DSC]

Since MikeC has the 5900 Ultra, why not ask him for some screenshots? :o

[Uttar]

Hmm, that's an idea...
But the screenshots should be 100% similar on a NV30, so I don't see the point in asking him that, frankly :) So could NV3x users give us some screenshots please?

BTW, could people with a NV30/NV31/NV34 benchmark this to see how much of an improvement the NV35 is in the FP32 shader domain?


Uttar

[Bopple]

Waiting for reply still...
If you calculate in FP32 mode, whole scene gets IQ boost upto FP32 rendered 'eyes' level, regardless of original data?

PS) You could have resolved those turmoils we had about FP matters quick/soon if only released this earlier.

[Uttar]

Originally posted by Bopple
Waiting for reply still...
If you calculate in FP32 mode, whole scene gets IQ boost upto FP32 rendered 'eyes' level, regardless of original data?

Okay, let me try to explain this...
The pixels get their data from the interpolated results of the three vertices forming a triangle.

The Vertex Shader *always* work in FP32.
That means the original data is *ALWAYS* FP32.
I think the NV30 it can be FP16, but I'm not sure about that. If so, then it might be possible to increase accuracy even further, although that would obviously be 100% invsible.


Uttar

[Bopple]

Thanks. That clears it up.

[GlowStick]

Originally posted by DSC
Since MikeC has the 5900 Ultra, why not ask him for some screenshots? :o

Sometimes review sites dont get to keep the hardware : (
but im not sure if he still has it or not : O

[jAkUp]

ok here is my results with a nv30:

running dawn ultra in quality mode in nv control panel, no aa or af... these are based off of sunning mode in dawn. 1024x768...
i dont have the fp16 so im unable to try that one...

regular shaders- 19fps
fp32- 15fps

[Nv40]

Originally posted by Grrrpoop
And despite Dawn not running in FP32 mode as standard, it still looks superb, which to me says FP32 is just bragging rights and not currently essential. I'd like to see some screenies of FP16 Dawn vs FP32 Dawn :)

I think ATI hit the sweet spot with FP24 tho. [/B]

errrrr...
THere are already game developers who thinks diferent than you ,
in fact they suggest that for some ultra-highquality photorealistic images even 96/128bits will not be enough.
you need higher precision specially for highly reflective surfaces ,
and for very high resolutions textures. DAwn demo doesnt have high reflective surfaces ,neither have more intructions than the limits of R3xx hardware.

a nice demo to measures the diferences will be TIMEMACHINE demo
of Nvidia , but it goes beyond the limits of R300 hardware in PS instructions
counts and you will need there higher precision ,for very nice reflections . :)

http://www.3dvelocity.com/reviews/nv30/nv30_3.htm

[DSC]

Oh man, Nvidia is really in trouble. 15FPS on a 5800Ultra, 18FPS on a 5900Ultra....

jakUp, can you please try it with AA and AF on?

Uttar, was the 5900Ultra's scores with AA/AF on or off?

[Clockwork]

Originally posted by Nv40
[B]errrrr...
THere are already game developers who thinks diferent than you ,
in fact they suggest that for some ultra-highquality photorealistic images even 96/128bits will not be enough.
you need higher precision specially for highly reflective surfaces ,
and for very high resolutions textures. DAwn demo doesnt have high reflective surfaces ,neither have more intructions than the limits of R3xx hardware.

a nice demo to measures the diferences will be TIMEMACHINE demo
of Nvidia , but it goes beyond the limits of R300 hardware in PS instructions
counts and you will need there higher precision ,for very nice reflections . :)


Nv40, why do you ignore the fact that:

A)nv3x is very slow at pixel shader operations (despite being capable of long instruction counts)


B)FP32 (128-bit precision) brings nv3x cards to their knees

:confused:

:rolleyes:

[jAkUp]

ok here is some pictures...
the normal shaders on top, fp32 on the bottom
hard for me to notie any differences, not to mention the screenies are compressed to get them under 100k

http://www.nvnews.net/vbulletin/attachment.php?s=&postid=129275
http://www.nvnews.net/vbulletin/attachment.php?s=&postid=129287

[Grrrpoop]

Originally posted by Nv40
errrrr...
THere are already game developers who thinks diferent than you ,
in fact they suggest that for some ultra-highquality photorealistic images even 96/128bits will not be enough.
you need higher precision specially for highly reflective surfaces ,
and for very high resolutions textures. DAwn demo doesnt have high reflective surfaces ,neither have more intructions than the limits of R3xx hardware.
..I think you misunderstood my post.

I'm talking about for TODAYS applications, whether that be games, benchmarks or demos.

FP24 is practical, FP32 so far does not seem practical, even in nVidia's own demo coded specifically for its own hardware. And that with a static pre-rendered background.

Future hardware from nVidia will no doubt run with FP32 more acceptably, but right now it's a bit sluggish, whilst ATI runs FP24 quite happily.

I eagerly await your next ill-conceived rebuttal :)

jAkUp: thnx for the pix, shame they're not in the same position but I know it's hard to do ;)

[GlowStick]

I have to say, the shaders dont look much diffrent.

[jAkUp]

edit:
ok i changed the above images to try to make dawn a little closer in position.. its not perfect now... but almost there:D

[Hellbinder]

To balance that statement out tho, the same can be said for the R350, with the exception that it actually performs very well at FP24 all the time..

Actually this is True... but.. Id like to point out that With the F-Buffer the Radeon 9800pro smoothly runs shader routines Hundreds of instructions in length, already Demoing a 500 instruction shader running over 30 FPS.

So yes if you actually totally maxed out both cards it would be at a slide show pace. However, I dotn think the Nv35 can even run a 500 instruction shader at anything evel close to 30 FPS.

[AnteP]

Originally posted by Uttar
Hmm, that's an idea...
But the screenshots should be 100% similar on a NV30, so I don't see the point in asking him that, frankly :) So could NV3x users give us some screenshots please?

BTW, could people with a NV30/NV31/NV34 benchmark this to see how much of an improvement the NV35 is in the FP32 shader domain?


Uttar

5800 @ 5800 Ultra
Normal: 30 fps
FP32: 17 fps

BTW when nVidia showed me Dawn at Comdex earlier this year they told me that it was capped at 30 fps so perhaps the difference is MUCH larger than what we are seeing.
I tried looking around in the config files but I didn't find any switch for any fps cap if there is such a cap that is.