Aniso Speed Improvement By The End Of The Week!!

[alexsok]

A couple of things:

1) Maybe that's the reason NVIDIA tried to stop the spreading of these drivers and remove them where they leaked, cause their speed doesn't match the speed improvement NVIDIA anticipates with the offical release.

2) As a member of a russian board (i'm a native russian), where Unwinder usually spills the beans on his latest discoveries (for anyone who doesn't know, Unwinder is the guy responsible for Riva Tuner, the best tweaking progie there is & among the best guys in the drivers department. hell, he could work for Nvidia!

Anyway, a post made by him on the board indicates that 10% improvement in anistropic filtering perfomance will occur by the end of this week, when an update will be released for Riva Tuner, which includes a little patch the guy made, that provides up to 10% improvement in anistropic filtering perfomance.

Also, he didn't see any improvement in the 30.30 set of drivers in this regard, except one small improvement, which is rarely used in testing...

P.S
The latest release of Riva Tuner can be downloaded here:
http://www.nvworld.ru/downloads/rivatuner.zip

P.S.S
To avoid confusion, this release does NOT include the patch that improves the anistropic filtering perfomance by 10%!

That should be ready by the end of this week.

Note: Obviously, in order to use the patch in the new version, an option would have to be enabled within Riva Tuner.
===============================================
O.k, I received a reply on my email from Unwinder!

Here it is:

Hi Dmitry

> Hey, sorry I can't write in Russiah, but i just don't have Russian fonts installed, so i hope it's o.k that i write in english

That's not a problem at all

> 1) Will the increase occur in Direct 3D or in OpenGL?

The patches will be able to boost OGL performance, whilst new D3D options will allow to boost D3D by disabling AF for specified texturing stages.

> 2) Where will it occur? On GF4/GF3/GF2/GF boards or only GF4 boards?

OGL patches affect all the GF boards, but new RivaTuner's D3D options are NV25 specific.

> 3) What improvements do u anticipate in this regard to be implemented in the next offical release of Nvidia detonators (probably 3xxx series)?

I hope that NV will implement more sophisticated AF optimization algorithm in the D3D driver, so the driver will be able to reject some textures for specified texturing stages instead of plain disabling AF for whole stage.

> 4) What increase will it be? 10%? 20%?

It's an application dependent issue. Average boost in OGL can be about 10% with no noticable IQ decrease, D3D boost can be much higher, but IQ decrease is much higher too).

WBW, Unwinder

 

[Sharkfood]

Well, the statement of "with no noticable IQ decrease" is obviously a subjective statement and it will be interesting to see what (if any) adverse effects to anisotropic filtering quality may or may not be experienced with more tweaks/tuning to AF.

I've already said it several times here- AF on the GF4 + current drivers is substantially poorer in quality when compared to AF + GF3 + 12.xx drivers in Direct3D. A simple game of Need For Speed-Porsche Unleashed is a good indicator of this, with it's gravel and road textures (Schwarzwald is a good example) being a flicker fest of shimmering and moire, whereas this was at least partially reduced in the past.

I've also several pictures showing the 4x->8x "transition" that occurs with the 29.42s on the GF4 (all in D3D to remove the confusion some are trying to add with Q and P modes in OGL). Quality differences are next to NILL and the only real markable difference is an extremely small difference in performance, with little yield to improved texturing or reductions in aliasing.

I'd happily *pay* an additional 10%-20% performance if we could simply get a clean, functional, and value added AF from the newest chunk of silicon. For now, I still prefer to play most my games on the GF3 or 8500 for reductions in aliasing, and only keep around my Ti4600 for it's excellent/flawless video capture and a daring hope that the fanbase will place less emphasis on fancy AF benchmark graphs and more on the delivered output quality/implementation.

 

[Pete]

Wowie, that extra 10% should have ATi running for the hills.

 

[pcchen]

I've already said it several times here- AF on the GF4 + current drivers is substantially poorer in quality when compared to AF + GF3 + 12.xx drivers in Direct3D. A simple game of Need For Speed-Porsche Unleashed is a good indicator of this, with it's gravel and road textures (Schwarzwald is a good example) being a flicker fest of shimmering and moire, whereas this was at least partially reduced in the past.

This game is not very representative, because it has very aggressive LOD bias setting for most of its roads.

 

[ConeK]

I was thinking the same. LOD bias can be altered easily with Riva tuner or with some other tweaker. There is also an article here whitch has some nice details of LOD biasing.

 

[Sharkfood]

This game is not very representative, because it has very aggressive LOD bias setting for most of its roads.

Alternatively, it is very representative as it is dramatic enough to be seen even in screenshots. Small bumps in LOD bias and increased texture filtering are usually only visible in motion and indeed this is the case with most everything in Direct3D.

The point is, 8x + GF3 + 12.90 compared to 8x + GF4 + 29.xx is a perverse, unusual and very visible difference/degradation in quality. Although it is *very* visible in other games, it's not subtle at all in NFS-PU when LOD Bias is carefully, painstakingly balanced and serves as a good example for illustrating the degradation. A user shouldnt be forced to tank LOD Bias to the point of blurring the scene dramatically to compensate for a diminished delivery of AF... which is the whole point.

And with the current public concern to AF performance, it seems to be mostly blind to AF quality.. from which a very real degradation has occurred, likely in favor of getting benchmark numbers up.

I hope this is a temporary trend and look forward to checking out the new drivers. If they are able to restore the IQ, even if it means at the cost of a reduction in performance, I'll be a very happy customer. If it's another shot of IQ disparity with slightly higher 3DMarks at 4x/8x... lets just say the trend is getting stronger.

 

[Nagorak]

Looks like they're going to need it, LOL!

 

[pcchen]

Alternatively, it is very representative as it is dramatic enough to be seen even in screenshots. Small bumps in LOD bias and increased texture filtering are usually only visible in motion and indeed this is the case with most everything in Direct3D.

Well, I don't think so. Aggressive LOD bias settings make texture aliasing. There is no good way to avoid this without using excess texture sampling. If you do anisotropic filtering in an "optimal way" (w.r.t MIP map LOD selection) you are not going to kill texture aliasing caused by aggressive LOD bias. That's why I think this game is not representive.

The key point is, most games should not adjust LOD bias at all. New games should rely on anisotropic filtering, not LOD bias. I don't know whether the newer drivers are more optimal than older drivers. However, if there are very little differences for normal LOD bias settings between the two driver sets, I won't say it is a "quality compromise."

 

[Sharkfood]

pcchen-

I totally agree with your concepts, but it isn't NFS-PU or LOD bias that is the root of the problem.

I guess you could condemn NFS-PU as the problem stems from it's grainy, single pixel detail textures, along with it's gridlike brick and building textures. These are the types of textures that show through aliasing quite well. But this game isnt alone- games like NOLF, Jedi2, Tribes1/2 and others also have the same kinds of textures that are problem children.

It isn't overly aggressive LOD either as lob bias adjustment works fine and dandy here... it just has to be tanked through the floor- to the point of making many textures excessively blurry to see.. this isnt the case on the GF3 or Radeon... you can still maintain an almost flicker/shimmer free output with still enough LOD Bias to see texture detail. It's a change in implementation of anisotropic filtering and possibly some small changes in mipmapping implementation that shows through. It's a good title as it can be illustrated in screenshots, but the same problem can also be witnessed with just about any other game but less so with screenshots. You have to 'be there' playing it in motion to pick it out, and its obvious/noticeable on a decent monitor.

 

[pcchen]

I don't quite understand what you mean... The problem with NFSU is not the texture type it used, it's the LOD bias settings. Its aggressive LOD bias generates texture aliasing even without anisotropic filtering, and anisotropic filtering is not supposed to make it better. Actually, anisotropic filtering generates more aliasing in most cases. That's why I said NFSU is not a good example.

I don't know about other games you mentioned; I have not played any of them on a GF4. I assumes that they don't use aggressive LOD bias settings like NFSU. If so and they still expose similar effects, NVIDIA should revise their current anisotropic filtering algorithm.

 

[Sharkfood]

That's the thing. It's NOT aggressive LOD Bias settings in NFS-PU. hehe. It's flat out texture aliasing, nothing more, nothing less.

Sure, LOD Bias can be softened to the extreme result of reducing the texture aliasing (as is the case with any game)... but the level you have to reduce this to fix the aliasing on the GF4 is absurd and not consistent with the 8500 and GF3. That is the point.

If you take a baseline IQ- with no AF, no AA, etc.etc. Get a comparable baseline of this across the platforms, you have a certain degree of texture aliasing. Throw on AF, say 8x level on all three products... texture aliasing is reduced/removed to a satisfactory amount, but it is still visibly apparent on the GF4 only. On the GF4 with 29.xx drivers, I am having to then dump LOD Bias at the +.75 to +1.50 level if I wish to remove the shimmering/moire that is completely gone on the other products. And if license plates, texture billboards and other textures are compared for visual sharpness, it is not simply an overly aggressive LOD Bias that is causing the problem. The texture detail sharpness or LOD is nearly identical with the cards, just the effectivity of the AF is greatly reduced in comparison. I hope that explains it better.

 

[Fred]

All the same, perhaps you're point would be better served with another example. Say from Serious Sam in motion, there's some scenes in the game with heavy texture aliasing.

 

[pcchen]

That's the thing. It's NOT aggressive LOD Bias settings in NFS-PU. hehe. It's flat out texture aliasing, nothing more, nothing less.

Well, I am not arguing about whether GF4 has good AF or not. I just said that the road of NFSU is not a good example for comparing AF quality. That's all. Other games you mentioned are perhaps much better examples, as I said in an earlier post.

The point I was making is, comparing AF quality based on textured triangles with very agrgressive LOD bias settings, such as the road of NFSU (it has -2 or even smaller LOD bias), is not appropriate. If you force the LOD bias to +2 or more in the driver panel, you can reduce the texture aliasing on the road, but also make other textures more blurry.

If I have some time I'll modify a program I wrote to be able to record a movie, for comparing texture aliasing effects when AF is enabled on different cards. That would be a clear evidence for the possibly inappropriate implementation of the AF kernel on a GF4.