Radeon 8500 Aniso vs Geforce 4 Aniso

[KimB]

Because it's obvious and I have a knack for stating the obvious?

 

[Doomtrooper]

I just have to say that the Radeon shot is full of texture aliasing. This texture aliasing will be far more visible when in motion.

Personally, I can't stand texture aliasing in the least.

The question is: Why is there lots of texture aliasing in the ATI shot?

There are two possibilities that I can think of:
1. ATI chose very aggressive LOD for the textures.
2. ATI's aniso algorithm doesn't use enough texture samples.

Given the very small performance hit of ATI's aniso, I suspect number 2 is the case. My main evidence for this is that even with 16x aniso, the Radeon 8500 has very little performance hit.

Besides, I'm not sure that just having too aggressive LOD would cause the jagged look that the Radeon 8500 shots seem to be showing.

ATI has a higher (lower) default LOD and has been proven MANY times..I choose #1 and as a OWNER of a 8500 I can say my choice is the correct one.. I wonder how many frames a 8500 could pick up if ATI allowed LOD adjustment ....

 

[Doomtrooper]

Because it's obvious and I have a knack for stating the obvious?

No, you have a Knack of stating what is obvious to whatever supports your opinion...

If you can't see in the center circle I marked that ATI's ansitropic is going back much further without blurring then you don't notice the OBVIOUS

 

[KimB]

If you can't see in the center circle I marked that ATI's ansitropic is going back much further without blurring then you don't notice the OBVIOUS

If ATI could do that without lots of texture aliasing, I'd call it a victory. As it is, what's to keep me from thinking that I couldn't get the same 'anisotropic going back much further' by just adjusting the LOD?

 

[Doomtrooper]

If you can't see in the center circle I marked that ATI's ansitropic is going back much further without blurring then you don't notice the OBVIOUS

If ATI could do that without lots of texture aliasing, I'd call it a victory. As it is, what's to keep me from thinking that I couldn't get the same 'anisotropic going back much further' by just adjusting the LOD?

Well there is not LOTS of texture aliasing in the 1st place...there is some, yet on my old Geforce 3 there was some too.
I had to use the in game control panel and drop the LOD on both Serious Sam games on both cards..although I had only the demo of SS2 before the Geforce 3 was sold.
LOD is not giving that effect as the filtering being trilinear, or Ansitropic will only filter so far, and the question on this thread is IMO and what Dave is asking..is there more than 10% of ansitropic filtering being used here by ATI.

 

[KimB]

Well there is not LOTS of texture aliasing in the 1st place...there is some, yet on my old Geforce 3 there was some too.

Yes, there is some on my GeForce as well, unfortunately, and it really annoys me...and there's no way you could convince me that the 8500 has better image quality if it obviously has more texture aliasing.

Btw, here's a couple of shots I took for comparison. Tell me which you think looks better:

http://169.237.254.233/img0.jpg
http://169.237.254.233/img1.jpg

(You may have to load them up in two different windows and swap between them to see the difference...).

 

[Ailuros]

Obviously, the rasterization engines should adapt their texture filtering and AA sampling to those areas which need them most, and they should endeavor to be "correct" most of the time and avoid artifacts by making the logic smart.

The man speaks to my heart

 

[Doomtrooper]

Well there is not LOTS of texture aliasing in the 1st place...there is some, yet on my old Geforce 3 there was some too.

Yes, there is some on my GeForce as well, unfortunately, and it really annoys me...and there's no way you could convince me that the 8500 has better image quality if it obviously has more texture aliasing.

Btw, here's a couple of shots I took for comparison. Tell me which you think looks better:

http://169.237.254.233/img0.jpg
http://169.237.254.233/img1.jpg

(You may have to load them up in two different windows and swap between them to see the difference...).

Does the screenshot show more than 10% of ansitropic filtering being used, the question that Dave B. has asked, no offense I don't need screen shots of Unreal Tournament to prove anything..I have put in over 40 hours in one week playing matches.

 

[Pete]

The 4600 shots do look almost antialiased--they seem a bit softer than the 8500 shot. The 8500 seems to have more detail, but I'm not sure if that's a good thing (and I haven't seen both cards in motion, which is key). The difference between the two cards is almost akin to setting the Sharpness on a TV up too high--seemingly more shimmering on the 8500. The floor tiles look more jagged on the 8500, but also slightly more detailed--they seem to "pop out" a bit more.

The only difference I see between the bi and tri 4600 shots is the latter shows more "filtered"/smoothed textures on the colorful tops of the left and right temples. That, and the green band around the temple, about 1/4 of the way down (near the top of the second "box" from the top), looks similar on the 8500 shot and the tri 4600 shot, but different on the bi 4600 one--I can't tell if the latter has more or less accurate detail. In fact, this whole paragraph may be moot because it seems you took the 8500 and tri 4600 shots from the same place, and the bi 4600 is a fraction of a step forward.

As for whether this is correct or superior aniso, I can't say, as I don't actually know what aniso is supposed to do. I just know that it tends to sharpen textures. As I don't know what aniso is, I'm not sure I can say whether ATi is aniso filtering more than 10% of the image. Their image certainly looks like it has more raw texture detail, but I prefer the 4600's more "blurred" look, particularly on the wall textures under the two side arches. The jagginess is also most evident on the floor adjacent to the right arch (perpendicular to the low wall under it).

The 8500 seems to have more accurate textures, though, as you can see more detail in the walkway all the way up to the door in the distance, and one texture in the temple ahead actually shows some detail that's missing from the 4600 (the red stripe about halfway up the first arch ahead of you shows a beige line at the lower edge that's not visible in the 4600 shots).

My question is whether the 4600 is "filtering" more than the 8500--whether the smoother image I see on the 4600 is actually aniso filtering, and the rougher yet more detailed image on the 8500 is actually just a more extreme LOD setting with less filtering.

I'd also like to know why the grass is missing in the 8500 shot.

And I'd like to know why you chose this shot for the comparison--it seems to me there'd be better locations for it, with more, and better lit, vertical textures, and preferably not at right angles to us (meaning, turn the guy 23 degrees right). You might try a flight sim, as well, as Rev suggested.

(Phew, that was a lot of uneducated analysis of what I consider a less-than-ideal example shot. )

 

[Oompa Loompa]

I'd also like to know why the grass is missing in the 8500 shot.

Grass removal is always an optimization. That's why lawnmowers are so popular.

 

[Pete]

Grass removal is always an optimization. That's why lawnmowers are so popular.

* rimshot *

That looks like greater than 10% filtering to me.

*crowd groans *

 

[KimB]

Does the screenshot show more than 10% of ansitropic filtering being used, the question that Dave B. has asked, no offense I don't need screen shots of Unreal Tournament to prove anything..I have put in over 40 hours in one week playing matches.

Of course it does. What the nVidia rep said was most certainly not true in that image.

However, I do have to say that more complex scenes may indeed show much closer to only 10% of the scene as using anisotropic. Of course, it's easily possible to make up a scene where 0% of the scene uses aniso on a Radeon...and I'm not talking about a scene that doesn't need aniso.

But, I feel that a more realistic estimate is that ATI's anisotropic filtering aniso's roughly 80% of the scene for simple FPS games (Quake3, UT, etc.), and closer to 50% for more comple FPS games (Morrowind, upcoming FPS games). Less than around 50% would only be in rather contrived scenarios...

The 10% figure was PR speak, what can I say? nVidia's been doing it for a long time, and I don't really care.

In an odd sort of way, though, he did speak the about the major problem of nVidia's aniso right now: It is applied to the whole scene. In fact, nVidia's aniso is applied to too much of the scene (at least in D3D). But, it still looks better than ATI's.

 

[PeterT]

#4 3D graphics is all about the art of cheating. Most people can't tell the subtle difference. While ATI's algorithm may be "less correct", no one is going to tell the difference, I bet, in the vast majority of game situations.

There's no way for ATI's algorithm to be "less correct", because there is no defined standard for doing anisotropic filtering. All that matters is image quality and performance, and IMHO ATI's implementation is not only clearly better in the latter but also arguably better in the former, at least in these shots.

Concerning ATI's implementation, there are only 3 widely known ways to do anistropic filtering: Ripmaps, summed-area tables and multiple sampling of the standard mipmap structure, as used by NV. As I said before, there's not a single correct algorithm, but different ones which have their own positive and negative aspects (The first two are quite memory intensive while the last eats lots of fillrate). I have no idea how ATI does their minification, but it seems like the smart way to do it.

 

[KimB]

There's no way for ATI's algorithm to be "less correct",

In my mind, there is only one way. That is if ATI's implementation adds visual artifacts to the image.

In fact, ATI's implementation does, as evidenced by those very screenshots. ATI's implementation adds noticeable texture aliasing, an effect which is far more visible, and far more annoying, when in motion.

Go back for a moment and look at the two comparison shots I posted. You'll probably come to the conclusion that one of them looks much sharper, and is therefore better-looking. The truth is that that shot has a massively-adjusted texture LOD (-1.0) that causes severe texture aliasing. It actually looks horrid in motion.

 

[Randell]

no really is that what you did, I never wiud have guessed

the ony person who can tell is if the 8500 shot aliases more than the gf4 shot due to greater LOD is Dave.

 

[noko]

Chalnoth,

Yes your second shot looks horrid and I too hate texture aliasing. SS SE with the LOD set aggressively (two notches in the sharp direction) looks terrible to me and I usually set it to Normal just so I don't get moire and texture aliasing. Without at least two shots with a slight angle change with the Radeon8500 and Ti4600 it would be somewhat hard to confirm servere texture aliasing at least for me.

Well if the Radeon8500 only requires to do 10% of the image while the Ti4600 has to do 100% of the image it is obvious that ATI is much smarter about how to do filtering. Why waste extra samples for filtering for surfaces that are directly facing you? That would be stupid as far as I can see it.

 

[Dave Baumann]

the ony person who can tell is if the 8500 shot aliases more than the gf4 shot due to greater LOD is Dave.

Nope, everyone can - the article is updated. I'll go through a number of replies in a second.

 

[jpprod]

There's no way for ATI's algorithm to be "less correct",

Actually there is, Radeon 8500's anisotropic quality on non-vertical/horizontal polygons is not optimal. See http://www.digit-life.com/articles/gf4/index6.html

 

[Doomtrooper]

Yes that shot shows what ATI's implementation has problems with, 45 degree planes and why the Reverend wanted a flight sim...yet when using the card this is not noticeable unless you REALLY look for it , if ATI can adjust the filtering to detect these 45 degree planes better as their Anistropic is adaptive in the R300..then there should be no flaws in their implementation.

 

[Dave Baumann]

OK, I’ll go over a few replies issues:

On your performance graph for the Radeon8500 the graph is labelled 16x while in the data section it is recorded as 8x. Which was it? I know it was the fillrate test in the graph, could you clarify.

Sorry, that was a typo. Both set are done using the highest aniso setting of each card to display their maximum anisotropic filtering quality, so that means 16x for Radeon 8500. I’ve updated that table but it might take a while for the web cache to clear to show it.

Also if you use the ExtremeGFX Add-on did you disable TRUFORM on the Radeon8500? Reason why I mention it is because the GF4 isn't able to do N-Patches which would penalize the Radeon8500 in your benchmark.

I believe it was disabled. However, the performance numbers are there to illustrate the performance difference of the particular implementations of filtering, not as card comparison (that’s why they are on different graphs).

And I'd like to know why you chose this shot for the comparison

Ease. I generally don’t play games looking for areas that would be useful for such comparisons. The SS:SE technology test is just an easy shot to take.

You'll probably come to the conclusion that one of them looks much sharper, and is therefore better-looking. The truth is that that shot has a massively-adjusted texture LOD (-1.0) that causes severe texture aliasing. It actually looks horrid in motion.

Chalnoth
- The shots I’ve used are both with default LOD (both game and drivers/tweakers) and I’ve updated the article to show the mipmap levels. You can see that both cards appear to have a similar point in which they change their first mipmap level so it would seem that there isn’t a significantly different LOD between the two implementations by default.