Someone please show me the difference in image quality on a ATI 9700 pro vs a GF4 ti?

[imtim83]

Ok here is a pic of unreal 2003 retail with my Leaktek GF4 ti 4400 overclocked to GF4 ti 4600 speeds. I am running unreal 2003 at 1600x1200 with all details on maximum. Please someone with a ATI 9700 pro on the same map and area with 6x or 4x FSAA and 16x AF on show me the difference with a pic. I don't see a difference in the pics on some reviews. I just can't see the difference. I want to but i can't.

Thanks


http://members.cox.net/imtim2/pic

[Raptorman]

I have a Radeon 9700. I will try and get you a pic before the night is over.

[thcdru2k]

it's not really the difference if the image quality, but the power of the r9700 to produce faster framerates of equal image quality. so in a sense, it looks better in motion. i doubt you'd be able to tell a difference from a screenie.

[SnakeEyes]

With the settings he mentioned, I'd have a hard time believing the 9700 isn't faster in UT2003. CPU limitations aside, that is. ;)

[StealthHawk]

can someone with a R9700Pro please benchmark Tenebrae Quake on a number of different resolutions?

[Chalnoth]

From what I've seen over at Beyond3D, the differences are as follows:

1. Radeon 9700 has an anisotropic implementation that's an improved version of what was done on the Radeon 8500. It still will look worse than the GeForce4 at some angles. To me, this makes it worse.

2. The Radeon 9700 has superior FSAA, for three primary reasons:

(a) More speed, due to higher memory bandwidth and other possible optimizations.

(b) Better sample patterns.

(c) Gamma-correct FSAA.

The benefit of the the higher speed is obvious, and the better sample patterns are almost as obvious for those in the known, but I believe both of those actually pale in comparison to the improvment offered for gamma-correct FSAA. Unfortunately, however, the amount by which the FSAA is gamma adjusted is currently not user-selectable, and will therefore be wrong for many people. Still, it is better than nothing.

A quick explanation of what gamma-correct FSAA does: When the voltage signal is sent to the monitor, doubling the voltage will not double the brightness. That is, it's not a linear response. As it turns out, most monitors have voltage response curves that just happen to fit in well with the human visual subsystem, so that while double the brightness isn't being produced, our eye sees it as that.

What this means is that when doing averages, straight averages are actually not the way to go. It is much better to do a reverse gamma adjust before averaging, and then gamma adjust the output. This won't affect the overall brightness, but will affect averages. While it is true that any average done (texure filtering, blending, etc.) should also do this, it really isn't feasible. But, FSAA can show the problems that result from lack of gamma correctness most when viewing wireframe images. Have you ever viewed an image in wireframe with FSAA enabled and noticed that, from a distance, the lines looked dotted? This should not happen with proper gamma.

In sum: The Radeon 9700 offers better FSAA and similar anisotropic (some say better aniso, I say worse) compared with the GeForce3/4 architecture.

[Bigus Dickus]

Originally posted by Chalnoth
From what I've seen over at Beyond3D, the differences are as follows:

1. Radeon 9700 has an anisotropic implementation that's an improved version of what was done on the Radeon 8500. It still will look worse than the GeForce4 at some angles. To me, this makes it worse.

Following that logic, the GF4 has an anisotropic filtering implementation that looks worse than the 9700 at some angles. Does that mean it makes it worse as well? Are they both worse, since they both look worse than the other in some cases? :confused:

I, like most other people, conclude that if implementation A looks better by 20% 80% of the time, and implementation B looks better by 5% 20% of the time, then impementation A is the better solution, and "looks better" overall. Strange how you reached the exact opposite conclusion.

[Chalnoth]

Originally posted by Bigus Dickus
[B]Following that logic, the GF4 has an anisotropic filtering implementation that looks worse than the 9700 at some angles. Does that mean it makes it worse as well? Are they both worse, since they both look worse than the other in some cases? :confused:

No, because the Radeon 9700's implementation is inconsistent. It may use an anisotropic degree of 16 on a horizontal surface, but will only use somewhere around an aniso degree of 4 or so on a surface at around the angle of 23 degrees. I claim that fluctuations like this are far more noticeable than the more consistent implementation available in the GeForce3/4 series.

In another manner of speaking, I believe that image quality is only as good as the lowest-quality part of the image. As a quick example, why increase image quality? The entire reason, for me, is so that I don't notice image quality problems (such as overly-blurry textures, aliasing, z-buffer errors, and so on). Since the Radeon 9700's worst-case anisotropic is in the region of 4-degree aniso, I don't consider it much better than 4-degree aniso. By contrast, the worst-case scenario for the GeForce3/4's using max aniso is still their maximum aniso (8-degree).

[Bigus Dickus]

So consistency is King then? I suppose 2x SSAA looks better to you then than 6x gamma-correct MSAA? No? What about those cases where an alpha texture might come into the screen? Well, that's the lowest quality part of the image. Using your rather clearly stated logic ("Since the Radeon 9700's worst-case anisotropic is in the region of 4-degree aniso, I don't consider it much better than 4-degree aniso"), you would conclude then that gamma-correct 6x MSAA isn't that much better than no AA? Hell, let's make it 2x performance Smoothvision just to give it some equivalent AA, which isn't all that spiffy IMO.

I mean, hell, sure it's jaggy as can be, but at least it's consistent about it, right? And since it's worst part of the image is better than the 9700's worst part, then it must be the better implementation, correct?

Apples to oranges you say? Why? Because it's a silly argument in this case... but why not in the case of AF?

You know, you had a point back when the 8500 lost all AF on 45 degree rotations, and had only bilinear so the moving mip-map line was actually noticeable (though personally I've never noticed it, not being a flight simmer). Now that trilinear blends the mip-map lines, and there is AF there (and I've seen the screenshots at B3D too, and in the worst case it's just a hair worse than the GF4 is there), I just don't see the argument.

I know you like to stick to your guns, and I know this has always been your trumpet call (lowest part of image defines IQ of entire image... at least where AF is concerned), but perhaps you should do a bit of deep thinking about this. If the AA example doesn't illustrate why, then perhaps we can find another example.

Again, I've seen the SS's at B3D. The 22 degree rotations were nearly identical on the GF4 and 9700 (much less difference than between the 4x and 8x mode of the 9700... though perhaps not between the 4x and 8x of the GF4 as it doesn't seem to do a whole lot more in 8x mode). On the other hand, on non rotated surfaces the difference isn't nearly identical, but very obvious.

Perhaps you're just arguing that a change in IQ, from whatever to whatever, is what bothers you? Then, if the 9700 changed from 32x to 16x (hypothetically) you would still conclude a straight 8x AF is better IQ? What if it chaged from 16x to 15x (not a real possibility, but used for rhetorical value)... would it be worse by virtue of it still changing in IQ, or would it be a "small enough" change to pass your threshold? You can see what I'm getting at here. You've defined your IQ criteria as "lowest quality part of the image" which is obviously silly, or perhaps you are implicitly defining it as being "the most consistent," since you did mention that, without qualifying statements about the required level of consistency, restraints on minimum level of IQ even if consistency is maintained (since you can run with no AA and no AF and have perfect consistency).

It's absurd, plain and simple, but to each his own I suppose.

[Chalnoth]

Originally posted by Bigus Dickus
[B]So consistency is King then? I suppose 2x SSAA looks better to you then than 6x gamma-correct MSAA? No? What about those cases where an alpha texture might come into the screen? Well, that's the lowest quality part of the image. Using your rather clearly stated logic ("Since the Radeon 9700's worst-case anisotropic is in the region of 4-degree aniso, I don't consider it much better than 4-degree aniso"), you would conclude then that gamma-correct 6x MSAA isn't that much better than no AA? Hell, let's make it 2x performance Smoothvision just to give it some equivalent AA, which isn't all that spiffy IMO.

I've posted on this again and again, and the answer is simple: the alpha test/MSAA problem is solvable through programming. The anisotropic problem is not.

And the other thing is simply that both the GeForce3/4 line and the Radeon 9700 use MSAA, so the point is meaningless here, whichever your stance.

Perhaps you're just arguing that a change in IQ, from whatever to whatever, is what bothers you? Then, if the 9700 changed from 32x to 16x (hypothetically) you would still conclude a straight 8x AF is better IQ?

Meaningless. Again, refer to my previous argument. I would consider such a technique little to no better than a more comprehensive 16x anisotropic implementation.

As a side note, I don't believe any current consumer-level video card supports non-power-of-two anisotropic degrees (Every pixel on the screen is either 1x, 2x, 4x, etc.).

Oh, and please stop with the pointless personal attacks. It seems that every argument I make is riddled with personal attacks in return. Try arguing the point for once. If you can't do that, then shut up.

[Bigus Dickus]

Originally posted by Chalnoth
I've posted on this again and again, and the answer is simple: the alpha test/MSAA problem is solvable through programming. The anisotropic problem is not.But it's not solved currently so my point is perfectly valid. I suppose then, that if an ATI hardware designer or driver delveloper were to reveal that the 9700's AF implementation could, in fact, be altered through programmability but they had simply not been sucessful at "correcting" the behavior, then your opinion would suddenly be reversed and the 9700 would have superior AF simply because it could be "fixed?" I find that unlikely. Until they actually solve the MSAA/alpha test problem, it's still a problem.

And the other thing is simply that both the GeForce3/4 line and the Radeon 9700 use MSAA, so the point is meaningless here, whichever your stance.Then you agree that the 8500's 2x performance smoothvision is of better quality than both the 9700's 6x MSAA and the GF4's 4x MSAA? Stop dodging the issue.

Oh, and please stop with the pointless personal attacks. It seems that every argument I make is riddled with personal attacks in return. Try arguing the point for once. If you can't do that, then shut up. :confused: :confused: :confused: I've done nothing but argue my point. I recall (and looking back at my posts in this thread) making no personal attacks against you at all. I think your opinion is absurd, but I never questioned your intelligence or anything remotely similar. Trying to cloud the discussion here instead of addressing the point yourself?

[Chalnoth]

Originally posted by Bigus Dickus Then you agree that the 8500's 2x performance smoothvision is of better quality than both the 9700's 6x MSAA and the GF4's 4x MSAA? Stop dodging the issue.

No. Not on the least. And I personally feel that both nVidia and ATI should be getting on game developers to fix this particular issue.

You should know by now that I prefer to argue more from a theoretical standpoint. From that standpoint, MSAA is most certainly a good thing (compared to SSAA) because it separates edge AA from texture AA. Because the two are fundamentally different, separating the two is necessary for the best performance/image quality ratio.

And as for alpha textures, the games that I play the most right now (Neverwinter Nights and Morrowind) use alpha blends. After I get UT2k3, and once the renderer source is released, I plan to see if I can't go ahead and use alpha blends there, too, just as I did with UT (Btw, alpha blends are used for some surfaces in UT2k3, just not all...though I'm willing to bet it'll be a fair amount more challenging to implement them in the remaining surfaces than it was in UT).

[Bigus Dickus]

I understand then. "Consistency" and "lowest quality part of the image" only matter in some circumstances. Apparently, it doesn't matter to you for MSAA because it doesn't happen that often. Oddly, rotated AF planes, although not very common either and not nearly as noticeable as a jagged alpha texture, does matter.

As I said... to each his own I guess.

[Chalnoth]

But rotated planes do happen, in each and every single game.

Alpha tests don't exist in every game.

And regardless of any visual quality benefits of returning to supersampling, it's not going to be done due to the incredible performance benefits from not doing it (though, in the future, selective supersampling may be done...but that will be up to game developers...and by selective, I mean per-primitive, not full-screen).

Update: I guess you could actually say that constistency only matters in some circumstances. Basically, it only matters when it's feasible. If game developers want to put alpha tests in their games, then they need to deal with the consequences (lower image quality on modern hardware...).

As a side note, regardless of your FSAA method, certain alpha-test situations will always show severe problems. Specifically, there will always be significant aliasing with an alpha test if the textures are minified too much (how much depends on the texture). An example of a texture that is particularly prone to this would be a chain link fence texture.

Regardless, I've said it before, and I'll say it again. Hardware companies are not going to switch back to full-screen supersampling, so it's pointless to even argue for them to do it. Therefore, optimizing image quality for multisampling hardware is up to the game developers, and it is certainly possible to do.

The anisotropic issue is totally different, as nVidia has shown that proper calculation of the anisotropic level is possible. They've been doing it since the original GeForce. This is an issue that can and most certainly should be solved by ATI.

[Chalnoth]

Oh, and one final thing. The worst-case scenarios for 2x SSAA and other forms of MSAA are the same, if you want to include alpha tests. Both will apply no FSAA in certain situations (MSAA due to how a game is made, SSAA at certain angles). You'd have to go up to at least 4x SSAA to make the appropriate comparison you were attempting to make.

[Bigus Dickus]

Originally posted by Chalnoth
But rotated planes do happen, in each and every single game.Of a few degrees perhaps. Most maps I play in UT and UT2k3 have rotations of just 10 degrees or so (more for small parts of UT2k3 simply because it has more complex surfaces, but these are mostly in models of statues and such, not repeating textures, though it does have more rotated plane terrains in some maps).

regardless of any visual quality benefits of returning to supersampling, it's not going to be done due to the incredible performance benefits from not doing itPerhaps the same can be said for ATI's AF implementation... regardless of visual quality negatives, it's not going to change due to incredible performance benefits from not changing (though, they did make a change from 8500 to 9700 for the better). I'm not saying this is the case, only that it might be, and you should recognize that it is a valid possibility. We haven't even touched on performance, since the discussion was on image quality. But, in some sense, that must eventually enter into the equation. It matters not if card A can do 32X MSAA if it's too slow for anything above 4X to be useful. In the case of the GF4, it has a large performance penalty, and in many cases 8X or even 4X is not useful. So, what is the better quality then... an implementation that is fast enough to run at 4X and always looks like 4X, or one fast enough to run in 16X and occasionally looks like 8X?

Hardware companies are not going to switch back to full-screen supersampling, so it's pointless to even argue for them to do it.For clarification, I'm not arguing in favor of SSAA. I agree with you completely that MSAA is the way to go. Separate polygon and texture AA, and use the most efficient way to address each. MSAA (or possibly FAA if further developed) is the best way to address edges/polygons, and I'm glad that ATI has made a move in that direction. I was simply using the problem with alpha tests as an example to investigate your logic.

nVidia has shown that proper calculation of the anisotropic level is possible. They've been doing it since the original GeForce. This is an issue that can and most certainly should be solved by ATI. You discuss this as if it's a "solution to a problem"... something ATI just hasn't been able to figure out yet. Have you considered the possibility that "fixing" this characteristic would have an accompanying couple of percent performance hit for each sampling degree? Again, I'm not saying this is the case, just that it may be. The ATI hardware developers seem to be rather bright, as the R300, overall, is an exceptional piece of hardware. Do you really think that they just can't figure out the damn equations to make AF consistent? No, I doubt it. It was likely a choice, either for transistor real-estate (which was probably rather tight on the R300) or for performance.

Oh, and one final thing. The worst-case scenarios for 2x SSAA and other forms of MSAA are the same, if you want to include alpha tests. Both will apply no FSAA in certain situations (MSAA due to how a game is made, SSAA at certain angles). You'd have to go up to at least 4x SSAA to make the appropriate comparison you were attempting to make.Fine, it was just an example. Use 4X performance smoothvision vs. 6X gamma-correct MSAA if you like. It's a hypothetical argument anyway.

[Chalnoth]

Originally posted by Bigus Dickus
[b]Perhaps the same can be said for ATI's AF implementation... regardless of visual quality negatives, it's not going to change due to incredible performance benefits from not changing (though, they did make a change from 8500 to 9700 for the better).

And I'm saying that there's no way this is the case. Just as you stated, in many games, the off-angle surfaces don't often occur (but they do occur sometimes in every game that has full freedom of view...regardless of the game's geometry). The performance drop of the GeForce3/4's anisotropic is most certainly not a direct result of its more comprehensive nature.

While it does, naturally, require more computational power to do proper anisotropic degree selection, and therefore more transistors, you'd think that ATI would be able to do it by now since nVidia's done it since the original GeForce.

Regardless, I've done some performance tests myself, and it appears that the primary problem with the anisotropic on the GeForce3/4, as it comes to performance, is that of the two textures possible per pixel pipeline per clock, if the first texture is set to use any degree of anisotropic, a second texture is not sent through (this means that there is a massive benefit from disabling texture stages 0 and 2).

So, you could say that it is an indirect result of the more comprehensive anisotropic implementation, as nVidia apparently did not put dedicated aniso-degree calculation hardware for both texture stages, but just one per pixel pipeline.

In the end, this means that if software wants to optimize for nVidia hardware (and they really shouldn't have to optimize in this fashion, but it is certainly possible), the base texture, and any other high-res texture, should be placed in the 1 and 3 texture stages (second and fourth, that is), where all textures that could stand not to use anisotropic should go in the 0 and 2 stages (first and third...). This will allow maximum anisotropic performance if aniso is then disabled for stages for which it is not necessary.

In the case of the GF4, it has a large performance penalty, and in many cases 8X or even 4X is not useful.

On my GeForce4, the only game that I don't use anisotropic on is the Tenebrae modification for Quake (Well, that and Diablo II due to a bug...but it wouldn't help Diablo II anyway). For every other game, I always run 8-degree anisotropic.

As far as performance, obviously the Radeon 9700 is the fastest available to date for games. That's not under dispute.

You discuss this as if it's a "solution to a problem"... something ATI just hasn't been able to figure out yet. Have you considered the possibility that "fixing" this characteristic would have an accompanying couple of percent performance hit for each sampling degree?

Well, obviously there would be some additional performance hit, but it won't be noticeable.

As one final note, if the NV30 does indeed ship as an 8x2 architecture, there's a good chance that the performance specs of its anisotropic implementation will be the same as the GeForce4's (subject to CPU limitations, of course). What I mean by this is that it will likely not be able to use the second texture per pixel pipeline if the first texture uses anisotropic. This may result in more of a performance hit for anisotropic, but should not decrease its performance below that of the Radeon 9700 (The NV30 should most certainly be faster due to later release, .13 micron technology, and nVidia's track record).

[Chalnoth]

Oh, and one final thing. The performance hit decreases as you increase the degree of anisotropic, due to the simple reason that less and less of the screen uses higher degrees of anisotropic.

This is another reason why the anisotropic degree selection algorithm that ATI uses just can't have a very noticeable impact on performance.

[ATI LoVeR 9700]

Chalnoth, you are probably one of the only 50 people on the planet that don't like ATi's Anisotropic method. Can you post me a screenshot of bad looking Aniso on the 9700 and then post the same pic with Aniso on the TI 4600. Show me the uglyness, please.

[Chalnoth]

Originally posted by ATI LoVeR 9700
Chalnoth, you are probably one of the only 50 people on the planet that don't like ATi's Anisotropic method. Can you post me a screenshot of bad looking Aniso on the 9700 and then post the same pic with Aniso on the TI 4600. Show me the uglyness, please.

It's been done before. I can't post a screenshot, as I don't currently own a Radeon 9700. I'm taking this information from what others have posted.

Here's what those screenshots have told me about the 9700's anisotropic (Statements assuming 16-degree anisotropic setting) :

For horizontal/vertical surfaces, the 9700 will use a max aniso of 16 degrees.

For 45-degree angled surfaces, the implementation is noticeably improved over the Radeon 8500's, but still falls short of 16-degree max aniso. It appears to be close to 8-degree max.

For about 23-degree angled surfaces, the implementation is at its worst, and is blurrier than a GeForce3/4 using 8-degree aniso. It appears close to about 4-degree max aniso.

If you yourself have a Radeon 9700, try the game Serious Sam: SE. There's a level with a rotating floor. If you stand at the door and watch the floor swing back and forth, you should notice how it gets blurrier and sharper as it moves.

[nin_fragile14]

1280x960, 6x AA and 16x quality aniso.

With all the jpeg compression garbage to go with it

http://lindset.codect.com/images/1034497339Shot000001.JPG

[nin_fragile14]

Also, performance aniso looks really ugly on some games. The textures are sharp, but there is a dark line that you can see a few inches away from your current position in the game that moves back as you move forward. (Quality aniso looks amazing though)

[Nv40]

The only thing, i dont like about ATI ways to do aniso
is that they selectively disable the degree of the quality
at some angles , from 16x to 4x at many angles.
while Nvidia in the other hand they always do 8x in their Geforce3/4 cards . its takes a huge performance ,but you have the option to choose to lower your settings to 4x/2x or choose performance tab +8x settings ,if your machine is not smooth at highest settings.

Notice the key word here is choose ,i think it is
alot better if ATi lets their customers to choose when they
want a part time 16x/4x aniso or when they want a
full time aniso like Nvidia have been doing always.
all-ways .

i can guaranteed you all that the radeon 16x anisoF will look
alot much better (at the expense of performance ofcourse)
if they do it at 16x at all times..

So i will not be so much surprised to see an Nv30 aniso at 12x
(with 80% surface covering -which i have heard somewhere..) looking much much better at all times than ATI's way of 16x aniso.

i predict that the incoming (128bits bus 4x1pipes) radeon9500 anisoF method will be even "cheaper" than the radeon9700pro method ,in the job it will do,something like 16x-8x at some angles and 2x/1x at others or who knows ,maybe
it will be the return of radeon8500 anisof "optimized method"
:rolleyes:

So the gamers will be happy again , more framerates ,with less
image quality , but who cares , you get exactly what you think you deserve. in the past ATi fans always felt comfortable
with the Radeon8500 AF (cough) method .


http://www.digit-life.com/articles/gf4/index6.html

radeon8500 do 16x anisoF at surface level angles and 0x aniso at non flat. and the radeon 9700 AnisoF is the same as
radeon8500 but this time the switch from 16x to 4x at some angles.. nice eh ?

and while the radeon9700 AnisoF is much much better that anything they have released and better at normal surfaces than the geforce3/4 ,is still questionable if the new
AnisoF method in the Radeon9700 (partime 16x+parttime4x
at some angles) is still better than the Nvidia anisoF8x fulltime
method in all kind of games where camera angles matters ,like water(aquamark)/race/flight/space simulations games , or combat simulations games like counterstrike /RaibowSix series and the hundreds of incoming rainbowsix clones that soldiers will be able to turn his head (peek right/left)at diferent angles to see in the wall corners if there is any enemy ,it will be a little
distracting to see the floor and walls magically lowering its image quality from 16x to 4x aniso in those games.. dont you think?


a side note..
-----------------------------

i dont trust too much in standar game benchmarks ,like most people , because timedemos like those made in quake3 engines
or unreal 1/2 engines tells you at the end only the average frame rate at x or y game . but they never tell you exactly
how smooth the game will run at all-times..
i noticed this when i bought my geforce4 the first time .
and launched quake3 timedemo. for my surprise the
performance with aniso8x was exactly the same in the
timedemo ,than my old Geforce3 ti500 !!!.
however in real game play ,it was Rock solid smooth playable quake3,medalof honor,wolftein at 1600x1200 and aniso8x ;
while my old Geforce3ti500 hardly was playable at those same
high quality settings but this time at 4x anisoF.

so i learned that while both cards GF3 and GF4 performance was
exactly the same , the geforce4 never took too much performance drops below playable smooth framerates
+/-35frames. right know as i have said many times i play all games at 1600x1200 at 8xaniso ,exept UnrealT2003 which as expected , it runs it at 1200x1024 ,high settings without aniso or aa in my Gf4ti4400 and atlonXp1900+ . ;)

[Bigus Dickus]

Reading that made my head hurt. Please excuse me, I'm going to go get an ibuprofen. If the dizzyness stops, I might return to address a couple of points.

[Chalnoth]

Originally posted by Nv40
Notice the key word here is [B]choose ,i think it is
alot better if ATi lets their customers to choose when they
want a part time 16x/4x aniso or when they want a
full time aniso like Nvidia have been doing always.
all-ways .

Not a possibility, NV40. It's a hardware implementation issue. Quite simply, it takes more transistors to do a comprehensive aniso technique.

i dont trust too much in standar game benchmarks ,like most people , because timedemos like those made in quake3 engines
or unreal 1/2 engines tells you at the end only the average frame rate at x or y game . but they never tell you exactly
how smooth the game will run at all-times..

Actually, UT2k3 does output spreadsheets that show exact framerates for each frame, from which you can easily compose framerate graphs. It would be nice if some websites would post benchmarks reflecting these results.