G80 indepth IQ reiew

[thilanliyan]

Originally posted by: BFG10K
If you're good Barney might visit later and show you what 2+2 is. Isn't that exciting?

lol...Barney....Haven't heard of him for a while. Does the blending of rings impact anything? Does a blended colour act as another ring?? (Just trying to learn btw)

 

[Gamingphreek]

The blending is merely the filtering between mip map levels. While AF increases the detail at each distance, the filtering (Trilinear/Bilinear) blend the the levels so that you dont immediately transition between the levels.

-Kevin

 

[thilanliyan]

Originally posted by: Gamingphreek
The blending is merely the filtering between mip map levels. While AF increases the detail at each distance, the filtering (Trilinear/Bilinear) blend the the levels so that you dont immediately transition between the levels.

-Kevin

So then is it still that the fewer rings you have the better filtering? How does having more rings translate to a difference in a game? Any screenshots comparing in game??

 

[Cookie Monster]

To my short knowledge on the AF colour rings, the different colour rings represent the amount of AF applied to that area. e.g 2xAF for the red colour ring. 16xAF doesnt necessary mean 16xAF is applied to the whole scene i think.

Unless im wrong here.

 

[Pete]

Maybe my answer here will help. The receding color rings represent increasing MIP-map levels, or smaller and smaller versions of the base texture. The further the colors are "pushed back," the more of the screen's textures will be sampled from the base, highest-res texture. You can see this in the split-screen 8x/16x AF pic on page 14 of AT's G80 review:* 16x "pushes" the MIP-maps further back into the scene (in this case, a checkerboard-textured tunnel) than 8x. The smoother, more gradual the blend b/w levels, the less noticable the MIP-map transition will be in-game. Bilinear is an abrupt transition, with no interpolation b/w MIP-map levels. Trilinear is a smooth transition. Most GPUs currently use a compromise (brilinear or trylinear, a mix b/w true bi and tri) for performance reasons. The same happens with AF: you may select 16x AF, but you're not getting 16x the samples for every single texture onscreen, as that'd slow rendering to a crawl. Everything in 3D is a compromise b/w quality and speed. G80 is another step in the right direction for texture filtering quality, no doubt helped by its immense raw power.

Keep in mind that, as Rys noted in the article that spawned yet another AT flamewar, that what you see in these AF testers may not represent exactly what you get in-game, but it's a reasonably accurate clue as to the quality you'll see in-game.

---------------
* BTW, I believe AT incorrectly uses the term "texture stage" to refer to MIP-map levels here. AFAIK, we're looking at a single texture stage--0 by default--and the colors represent ascending MIP-maps, not texture stages. IIRC, this was an issue with UT2k4 during the bad old days of the race to the bottom of IQ levels for the sake of benchmark speed, where only texture stage 0 would get AF and the rest (if 8 stages, then 1-7) would just get bilinear.

 

[Pete]

WRT the ongoing "discussion," I thought that Wikipedia AF article is based on this SGI article and so describes (a form of) RIP-mapping, which is different from how GPUs implement AF. RIP-mapping, like bilinear and trilinear filtering, seems to isotropically sample from anisotropic MIP-maps. I thought GPUs anisotropically sample from the game-supplied (default/normal/game-supplied/and I'm presuming isotropic/square) MIP-maps. Who's right and who's wrong?

(Easiest thing would be to ask ATI/NV, but maybe checking VRAM consumption with AF would be a clue? That's assuming SGI's AF MIP-maps take up more memory than the normal ones they replace, or that they don't replace them at all but accompany them. I'm still not sure if all game devs allow for auto-generated MIP-maps, which is why I'm assuming current AF is anisotropic sampling on isotropic MIP-maps.)

Greg, your (crop of B3D's) NV30 vs. R580 pic is interesting. Clearly, NV30 is less angle-dependent, specifically at 45 degrees from the major axes, and so superior in that respect. However, R580 has superior MIP-map transitions. NV30 has some funky dithering and evident banding in its color transitions, but I don't know if that's all brilinear or exactly how the dithering manifests itself in games.*

Mainly, I don't know why you stress the importance of those "extra" center colors (MIP-maps). I didn't realize they'd matter at that view distance, nor that not seeing them implies a GPU stops using MIP-maps beyond a certain distance. Is the extra touch of yellow the reason G80 reportedly shimmers less than R580? Perhaps not. Despite "missing" those extra MIP-maps (which, being lower res, should mean a blurrier image), G80 looks sharper than G71 in all of Rage3D's AF test cases.

OTOH, something's happening in GTR2 and UT2k4 that shows a tradeoff b/w clarity and shimmering.**

------
* I'm not even sure how that dithering manifests itself in motion; maybe texture shimmer (pixel popping)? The banding probably meant some brilinear funkiness.

** Read Ail's take here and his conversation with Ratchet starting here.

 

[BFG10K]

So then is it still that the fewer rings you have the better filtering?

Yes. For filtering quality what you're looking for:

• Having as few rings as posisble.
• As smooth transition as possible betwen the rings.
• The rings being as close to a circle as possible.
• The whole "rainbow" being back as far as possible.

The G80's AF is so good because it comes closer to all of those goals than any other consumer card.

How does having more rings translate to a difference in a game

Having more rings makes the textures in the scene blurrier because you're getting a reduction in AF and hence being forced to use more mip-maps to compensate.

To my short knowledge on the AF colour rings, the different colour rings represent the amount of AF applied to that area.

No, each colour ring represents a mip-map. The red ring is the first mip-map, green is the second, etc.

 

[Gstanfor]

G80 is very much like nv4x and g7x in that it is still very sensitive to LOD settings (which is what causes most of the AF problems that the lack of blending doesn't account for on nv4x / g7x). If you look around the forums, especially at posts by chrisray & ailuros you'll see g80 isn't immune from shimmering.

By contrast, you'll be hard pressed indeed to get nv3x to shimmer (its different AF algorithm doesn't use LOD the same way as later chips do). Hence my belief (stated earlier in the thread and queried by CM that g80's AF is more closely related to nv4x/g7x's than it is to nv3x/2x).

Nv3x uses Centred weighted Average filtering - everything following it uses manhattan filtering.

People should try running the AF app with different LOD settings and LOD clamp settings.

 

[Acanthus]

I havent read an article at B3D for quite some time because of their history with severe bias...

I dont read THG for the same reason.

Im not suprised at all that they would play down the advantages of G80.

 

[ginfest]

Originally posted by: Acanthus
I havent read an article at B3D for quite some time because of their history with severe bias...

I dont read THG for the same reason.

Im not suprised at all that they would play down the advantages of G80.

QFT-not surprising considering who runs that site now