Anistropic Debate

[Doomtrooper]

This posted on Raged3D and Nvnews, although not too indepth there is alot of screen shots to look at, and judging by this review the Geforce 4 is taking a massive 60%-70% performance hit with new engine games with 8X anistropic enabled.

I am well aware of the differences in quality (IMO not much different between the two) so in the future what do you think will be done to get a happy medium between ATI's and Nvidias approach.
Obviously Trilinear and Ansitropic is not the way to go, and some skeptics gripe about the mip map borders that can occasionally be seen on Ripmapping.

So where do we go next..


http://195.38.1.194/test/

 

[John Reynolds]

Please don't drag flame wars from fan sites over here.

 

[LeStoffer]

So where do we go next..

I would like both ATI, nVidia and [vendor] to give the user the option of using either Performance AF or Quality AF. Implement both for crying out loud!

Regards, LeStoffer

 

[Joe DeFuria]

Well, the "happy medium" to me is simple: someone should implement both rip-mapping and "true" anisotropic, and give the user the choice of which one to use.

I'm waiting for the day when some company with give complete control over the "filtering method" to the end user based on a control panel applet. Choices should include:

1) Bi-linear
2) Tri-linear. (With sub-options for "true" trilinear, or "fast" trilinear.) "Fast" tri-linear might be implemented differently by different companies. Some might use auto mip-map generation with proper blending, some might use mip-map dithering instead of blending, etc.

In addition to those options, the user could also apply "anisotropic" on top of the above:

a) "True" anisotropic, with desired tap level
b) "Rip-Mapping" (which would force bi-linear, AFAIK).

I really wish some hardware vendor would introduce something like that...

 

[Mephisto]

Doomtrooper, your conclusion is wrong. You can implement anisotropic filtering without a performance loss compared to trilinear if you spend enough time on engeneering and enough space on your silicon.

It's all a question of question of cache size, caching algorithms and TMU design. GF3/4 looses a lot of performance because it is designed to fetch only 8 texels/clock and it has small texture caches.

R200 does not loose that much performance because it has the better caching due to the different, more cache friendly but lower quality texture sampling algorithm as well as the smaller amount of data (just one mipmap to read from).

I'm pretty sure future parts will offer both, high quality anisotropic filtering with mipmap interpolation and small(er) performance hit.

 

[Doomtrooper]

Please don't drag flame wars from fan sites over here.

Excuse me Johnny Boy, how is talking about what needs to be done to improve anistropic performance and visuals a flame war.

Grow up.

 

[Doomtrooper]

Well, the "happy medium" to me is simple: someone should implement both rip-mapping and "true" anisotropic, and give the user the choice of which one to use.

I'm waiting for the day when some company with give complete control over the "filtering method" to the end user based on a control panel applet. Choices should include:

1) Bi-linear
2) Tri-linear. (With sub-options for "true" trilinear, or "fast" trilinear.) "Fast" tri-linear might be implemented differently by different companies. Some might use auto mip-map generation with proper blending, some might use mip-map dithering instead of blending, etc.

In addition to those options, the user could also apply "anisotropic" on top of the above:

a) "True" anisotropic, with desired tap level
b) "Rip-Mapping" (which would force bi-linear, AFAIK).

I really wish some hardware vendor would introduce something like that...

I like the idea of having that option too and a LOD slider like 3DFX did as adjusting the LOD will help with mip map borders with the bilinear option.

 

[Typedef Enum]

Personally, I like choice...and as such, I think the obvious 'happy medium' is to give the end user the choice between a Performance and Quality mode, much like ATI does with SmoothVision.

I think the main problem you have in getting such a thing from nVidia is the fact that they seem to strongly believe in doing things certain ways...and in this case, they seem to feel that there is a definite right vs. wrong way of implementing this feature...despite the advantage to the end user and the fact that there isn't a written-in-stone approach...but it's clear that they make a distinction between what they offer vs. ATI.

Honestly, if there were enough people barking down their tree wanting to have a certain feature, I wouldn't be surprised to see them do it...But I don't think they would do it otherwise.

As a point of reference, consider the 4xs AA mode...I asked a long time ago about the possibility of getting support in OpenGL, and was flat out told "No..."...Since the debut, one of the interviews (cant recall which) with an nVidia rep point-blank said that their engineers were working on the OGL implementation...I think they're responding to both people asking, "Why only D3D?" and also wanting to give people an additional reason to purchase such a card (in the cases where users already have, say, a GF3).

About 1.5 years ago, it was all about getting better FSAA performance...and I think that has been achieved (more or less)...a short time later, a little more emphasis was placed on Quality (IE SmoothVision and 4xs)...and now the trend clearly is to bring that up a couple of notches, to include anisotropic performance.

I will be really dissapointed if the true next-gen parts don't offer better aniso. performance AND better/higher quality AA.

 

[Doomtrooper]

Looking at the AA peformance of the Ti4600 I'd say FSAA peformance is maturing and all I see on future cards including the R300 is more attention spent on FSAA. I play with Anistropic on all the time and I personally feel it gives a better visuals effect than FSAA.
So I hope there is some work being done to improve anistropic too.

 

[Mephisto]

and in this case, they seem to feel that there is a definite right vs. wrong way of implementing this feature...despite the advantage to the end user and the fact that there isn't a written-in-stone approach...but it's clear that they make a distinction between what they offer vs. ATI.

IMO this is part of their marketing concept. They want make fansites believe that ATis way doing anisotropic filtering is wrong and Ati "cheats" somehow.

 

[LeStoffer]

As a point of reference, consider the 4xs AA mode...I asked a long time ago about the possibility of getting support in OpenGL, and was flat out told "No..."...Since the debut, one of the interviews (cant recall which) with an nVidia rep point-blank said that their engineers were working on the OGL implementation...

Good idea: You and others with connection to nVidia should start asking whether they are able to implement a quick and dirty AF or not. Just challenge them and I'm sure you will get some kind of response or another.

Regards, Lestoffer

 

[Fred]

IMO ATI's ripmapping is the wrong way to go. Nvidia's method looks better, and while it takes a speed hit at this time/generation, silicon shouldn't be wasted on an inferior image quality solution. That would be like if we wanted to go back to Ogss compared to quincunxx.

In fact, a nice gaussian filter would be even more ideal (and probably coming soon) and smarter adaptive anisotropic logic would be a more elegant solution.

 

[John Reynolds]

Excuse me Johnny Boy, how is talking about what needs to be done to improve anistropic performance and visuals a flame war.

Grow up.

You have been repeatedly warned by Wavey about name calling and personal attacks. I have never been disrespectful to you and I would appreciate that courtesy being returned.

This topic has been nothing but a giant flame war between the regulars at Rage3D and NVNews. Personally, if someone wanted to objectively discuss the differences between Nvidia and ATi hardware, I don't think they'd initiate the discussion by citing a 'review' that does nothing but bench at 16x12x32 to show what a "massive performance hit" one piece of hardware takes while downplaying the IQ issues of the other ("some skeptics").

You're welcome to post on these boards, but I'd like to see the fanboy noise levels turned down a notch or two. I honestly do believe that heavy-handed moderation is counter-productive, so I don't want to edit/delete your posts. Just please try to keep things on a more even keel on these boards.

 

[Typedef Enum]

"They want make fansites believe that ATis way doing anisotropic filtering is wrong and Ati "cheats" somehow."

I don't believe they perceive ATI to be cheating...it's just the simple fact that their method doesn't address the issue across the board. Clearly, they're not wrong in this regard.

Now, whether or not you think nVidia's implementation is worth the performance hit is another question. As I said, I would be OK with a perf. vs. quality choice, though I would probably tend to go with Quality the vast majority of the time.

 

[Mephisto]

Nvidia's method looks better, and while it takes a speed hit at this time/generation, silicon shouldn't be wasted on an inferior image quality solution.

According to this argumentation, Nvidia shoudn't have "wasted" silicon space for their inferior multisampling and should have used supersampling instead.

That would be like if we wanted to go back to Ogss compared to quincunxx.

Quincunx is the blur filter I think we all don't like.

What you probably mean is going back from a RG pattern to a OG pattern in general.

 

[Doomtrooper]

You have been repeatedly warned by Wavey about name calling and personal attacks. I have never been disrespectful to you and I would appreciate that courtesy being returned.

This topic has been nothing but a giant flame war between the regulars at Rage3D and NVNews. Personally, if someone wanted to objectively discuss the differences between Nvidia and ATi hardware, I don't think they'd initiate the discussion by citing a 'review' that does nothing but bench at 16x12x32 to show what a "massive performance hit" one piece of hardware takes while downplaying the IQ issues of the other ("some skeptics").

You're welcome to post on these boards, but I'd like to see the <bleep> noise levels turned down a notch or two. I honestly do believe that heavy-handed moderation is counter-productive, so I don't want to edit/delete your posts. Just please try to keep things on a more even keel on these boards.

Well what happens on Rage3D and Nvnews has nothing to do what I posted here. I titled it a debate as I think this is a important subject .
The review was important also because it shows what I was talking about with screen shots.
Your free to edit my posts because honestly its a sure sign of BIAS. I never posted any negetives about any product besides a massive performance hit mentioned, and for a $700 dollar video card where I live, if it can't Deliver MAX Anistropic at 1600 x 1200 there is a problem..dig.

There was some good comments on this thread too bad it was ruined by a moderator that maybe too hasty with the extreme power approach. Or didnt like the review

 

[Fred]

The difference Mephisto, is that you don't lose any image quality with multisampling compared to supersampling as long as you are using anisotropic filtering. (well assuming you neglect or precheck alpha textures)

An argument can be made, that Quincunx generally looks better than the equivalent sample ogss mode, although at a speed cost.

But sure, rgss compared to ogss works too, for the line of argument.

All i'm saying, is that it strikes of shortsightedness to go for a solution that is inferior in image quality, but with a slight speedup.

These engineering trade offs are also not made in a vacuum. Adding silicon for ripmapping say on an Nvidia part would be ... questionable. Is the added heat/power cost/clock reduction/die space lost really worth it?

 

[Sharkfood]

Flamewars only arise when uneven/biased comparisons of products are sadly attempted to be made. I don't think there is anyone here stupid enough to fall for those tactics.

Although I agree with "more is better" and that having tunable texture filtering, LOD Bias and other such control left for the user, I don't think the average website reviewer could handle this amount of complexity.

It would simply become a war of default settings, which is what we have today. Information on the web is like "facts in a vacuum" at current as none of the ratings systems or product score/evaluations out there really pay any attention to the end result- and if they do, it's done for reasons of agenda as they will do a mutli-page derogatory analysis in favor of one products weaknesses and completely glance over anything that might hint at the reverse. Selective information.

Adding more end user control will just be additional fuel for the selective presentation of facts.

How many reviews now have tons of pictures of anisotropy- rotated, zoomed, anim gifs and the like? And how many have the same attention put towards Antialiasing? It just seems such scrutiny and criticism over image quality, albeit NOT a bad thing, is very specific and completely blind with similarly excessive IQ differences.. which are conveniently skipped over and ignored. If an analysis is to be comprehensive, it should be just that... not what we have today.

What it's really coming down to is- what is good for the gamer and end consumer is not good for a company selling it's products. The more settings end user reachable, the more conservative the defaults will be set to and the more uneven comparisons will be... or more uneven with incredibly selective commentary and analysis.

So I don't know an answer to all of this, at least until the "cookie cutter" benchmark graphs and "9 out of 10" review process is made obsolete and the emphasis is returned on the final resultant image and non-selective information. The moment this occurs will be the moment videocard manufacturers can start putting emphasis on real value versus fictional website races.

It would be a bold leap for ATI, NVIDIA, Matrox or anyone else to put a comprehensive and fully adjustable set of parameters into their drivers... and one that would stand to reason to be exploited if past history is any indication.

 

[ben6]

EDITED: Sorry for any confusion, but I won't be posting anything regarding anisotropic filtering in this thread...


Ben

 

[Ante P]

here's a "nice" nvnews discussion

http://nvnews.net/forum/showthread.php?s=&threadid=12569

I'm Undac in there. Most of them seem to hate me, honestly I don't know why