Aniso becoming less useful?

[BoddoZerg]

In UT2003 screenshots we saw that anisotropic filtering inflicts an incredibly large hit compared to the relatively small hit from aniso on current games. At the time I figured it was a fluke with UT2003. However, after trying the "Tenebrae" quake1 mod, I'm starting to wonder if massive framerate hits from anisotropic filtering are going to become standard as more games use advanced lighting and bumpmapping?

Playing tenebrae on my AthlonXP2000/GF4Ti4400 system, I was getting nearly unplayable framerates even at 640*480 with 8x aniso... after turning aniso off, I could play at 1024*768 with perfectly smooth framerates.

Is there something about advanced lighting techniques that makes aniso extremely slow? Hell, 8x aniso is slower than 4x fsaa...

 

[Hyp-X]

GF4 cards has a serious problem with anisotropic filtering.
The 2nd texture unit is disabled when AF is enabled.
nVidia probably simplified that 2nd TMU so it's not capable of AF.

I wonder if GF3 cards could beat Gf4s in these games with AF.

 

[Nagorak]

GF4s really are dogs for Aniso. R8500 is actually faster than a Ti4600 when aniso is on.

 

[Reverend]

Perhaps RivaTuner may help with its "Performance Aniso" and "Quality Aniso" settings. The "Performance Aniso" really boosts performance (compared to "Quality Aniso" that is) with very little noticeable difference (again, compared to "Quality Aniso")

 

[Galilee]

Yes that feature in Rivatuner is great:
Did a quick performancetest in Q3A. 1024 Max quality. (Q3A scales well when using aniso, compared to lets say SS).

No aniso : 178.2 fps

2X Quality: 153.7 fps (-13.75 %)
4X Quality: 127.7 fps (-22.34%)
8X Quality: 112.7 fps (-36.76%)

2X Performance: 167.4 fps (-6%)
4X Performance: 150.6 fps (-15.49%)
8X Performance: 137.5 fps (-22.83%)

I just noticed in Anandtechs review of R300, when he tested aniso in UT2003. The performance hit with Quality aniso were pretty similar to my numbers above (13.75....) a little bit higher. Performance were a little faster than Performance above.
Yes these are two different games, and yes one is OpenGL and one is D3D. Just saying that aniso is not for free when using Quality on ATI cards.

 

[alexsok]

If u set in Direct3D aniso level 8 and limit only the first stage to level 0, you'll see a very small perfomance decrease (and the gf4ti4600 becomes faster than Radeon 8500).

I think maybe NV28 will do something about that without using RivaTuner.

 

[Bjorn]

If u set in Direct3D aniso level 8 and limit only the first stage to level 0, you'll see a very small perfomance decrease (and the gf4ti4600 becomes faster than Radeon 8500).

And nothing happens with the image quality ?

 

[LeStoffer]

Another view on the topic, that I have been thinking about. As we move forward towards every pixel being rendered by a shader (and using per-pixel lightning) we won't be bound by the simple rendering metode like Phong or Gouraud shading.

This in itself could increase perceived detail - just have a look at any game that are using pixel shaders today. You won't be taking a texture map (or mip-map) and just render it simple as that in the future. You might still need AF in order to combat texture aliasing with a given pixel shader, but I don't have the knowledge to flly understand this issue.

 

[alexsok]

And nothing happens with the image quality ?

Well, there is one review coming up that deals with this subject in a little while and I all I've seen for now are only benchmark pics, nothing more.

I do think there will be IQ decrease (I limit Stage 0 to leave 0).

In 3DMark 2K1SE for instance, I got 10000+ without Aniso and 8500+ with Aniso 8 and the limit I set (on gf4ti4600).

When the review comes out, you'll see exactly how small/big the decrease of IQ is.

On a related note, I yesteday tried to play MotoGP (it uses PS &VS and all the latest features) with Aniso 8 & 4xS AA.

The image quality was great and I almost noticed no perfomance difference from no AA & no Aniso!

 

[MikeC]

And nothing happens with the image quality ?

Well, there is one review coming up that deals with this subject in a little while and I all I've seen for now are only benchmark pics, nothing more.

Digit-Life did a piece on the different anisotropic options available in RivaTuner. Nice selection of image quality comparisons as well.

http://www.digit-life.com/articles2/gf4/gf4ti4200-3.html

 

[alexsok]

Digit-Life did a piece on the different anisotropic options available in RivaTuner. Nice selection of image quality comparisons as well.

http://www.digit-life.com/articles2/gf4/gf4ti4200-3.html

Yes, but that article was with gf4ti4200 while the article that is going to be published soon is with gf4ti4600 and a bit different settings.

 

[Galilee]

After reading that review and from my understandig of the stage-settings I think they have misunderstood the setting.

They got around 160fps in Game1 Low detail without aniso(same as me, same card). When I tweaked the aniso with 8X stage0, 0X stage1-3 I got 137fps, they got somewhere around 100fps with optimized stages.

I just don't understand their settings. They use 8X but optimize stage 0 with 2X, ?? wtf. Leave stage 0 alone, and optimize on stage 1-3.

 

[BoddoZerg]

What are stage 0, 1, 2, and 3?

 

[Galilee]

From Rivatuners helpfiles:

 

[malcolm]

How many texture samples does anisotropic use?
It used to be in tabs now they say 2x 4x ...

 

[Xmas]

How many texture samples does anisotropic use?
It used to be in tabs now they say 2x 4x ...

If you enable 8xAF with trilinear filtering, you get at least two and at most eight triliear filtered samples, that's between 16 and 64 "taps" or "unfiltered texels". Bilinear cuts that number in half.

 

[tEd]

What are stage 0, 1, 2, and 3?

stage0 are the main textures so if you want anisotropic you better leave that stage unoptimized while stage1 are lightmaps where you can disable anisotropic and force bilinear to increase speed , for the other stages it depends on the game and how many texture layers they use but the most games use only stage0 and stage1.

 

[Joe DeFuria]

Hmmm...

[sarcasm]
It looks like nVidia is "cheating" here by not applying aniso to all texture stages. And ACK....disabling trilinear filtering? Surely, no self-respecting GeForce owner would enable such a thing, no matter what the performance boost. Can't nVidia engineer an anisotropic engine that is fast without having to resort to implementing all of these cheats and hacks? Man, that implementation sucks!

[/sarcasm]

 

[BoddoZerg]

[sarcasm]
wow, look at us, our sense of irony is so refined that we use SARCASM html tags!
[/sarcasm]

 

[OpenGL guy]

Hmmm...

[sarcasm]
It looks like nVidia is "cheating" here by not applying aniso to all texture stages. And ACK....disabling trilinear filtering? Surely, no self-respecting GeForce owner would enable such a thing, no matter what the performance boost. Can't nVidia engineer an anisotropic engine that is fast without having to resort to implementing all of these cheats and hacks? Man, that implementation sucks!

[/sarcasm]

So what happens if this treatment of anisotropic filtering gets put into the driver to improve performance?