Nvidia GF4 Aniso and FSAA

[Hellbinder]

A coupple of questions.

First for those of you in the know (becuase you own both).

Does the GF4's FSAA look as good as Smoothvision? or is it still blurry. Does the 4XS work well, does it improve IQ? and is compatible with all games?

Now, aniso.

The long fight between ATi and Nvidia fandom over which is better bla bla bla... I know that Nvidia says they are focusing on Aniso performance in their comming drivers. Is it possible to impliment a Aniso scheme simmilar to Ati's? Is this something that is strictly hardware? or could they code their drivers to have a special "mode" that is called fast Aniso or something. That would deliver far greater speed, but be somewhere inbetween Triliniar/Ansio the full screen. I am not sure if that is a pure hardware design function or drivers, or just how much of a blend of the two it is.

Thats it folks. any commentary?

(I am considering getting a Ti 4200 for my sisters fam)

 

[Sharkfood]

I own both and have posted on this subject many times.

No, the GF4's FSAA is not equivalent nor comparable to the 8500's Smoothvision. In general, their anisotropy has a much more similar effect compared to the 8500's SV in respect that it applies a benefit to the entire scene.

By enabling 2xAA on the GF4, edges are smoothed... and *very* effectively if I might add. It looks really nice, but does nothing to anything else in the scene. The overall scene smoothing you immediately notice with the 8500 isnt there.. it's just the edges, nothing else but.

On Quincunx- it's still a post filter blur and has adverse effects on the full scene in the form of reduced texture clarity, and not even like standard supersampling. Supersampling may have an overall smoothing on textures, but not a blur/smear mask like Quincunx does. Still, Quincunx isn't that bad at very high resolutions in flicker-fest titles or low detail textured games (Quake2 + Quincunx is actually pretty decent due to it's chromagnon textures).

On 4xS- it's everything it's been touted to be. It is limited to ONLY Direct3D games and I havent found anything it doesnt work with.. (yet). It's very clean but also just as expensive as SV on the 8500 in terms of performance. You get what you pay for in both cases- VERY clean AA. It's just the hit severely limits it to only a handful of titles.

Anisotropic filtering is just screwed right now on the GF4 + Direct3D and I've posted this several times. It was better in previous drivers on the GF3. Right now it does very, very little... does absolutely NOTHING between 4x and 8x in Direct3D, and carries a big hit. It's something NVIDIA is obviously working on since my GF3 + 12.90s has better visual quality with 2x,4x,8x in both OGL and D3D, yet throw on the 28.32s or 29.42s and it's demoted to a GF4 in terms of quality and performance. The fact that 4x->8x does nothing (along with all the registry key changes and God knows what undocumented registry exclusions or whatnot are hidden) tells you it's in progress as we speak, so the next "Det5's" or whatever should hopefully yield some relief here. For now, Direct3D has a weak 2x, pretty decent 4x, non-existent 8x (performance only, no IQ change).. and OpenGL has a decent 2x, decent 4x, next to impossible to see 8x... again strictly focusing ONLY on IQ and not performance.

If I can scrounge up enough space on my Tripod, I'll take some more Morrowind shots with both the 8500 and GF4 Ti4600.. running "raw" and with AF, AA and AF + AA.

I already have some GF4 shots you can look at (especially concerning performance and the cryptic 8x D3D AF) at:
http://shark_food.tripod.com/nfs/nfs.html

I'd be interested if other GF4 owners can play with 4x and 8x anisotropic and see if they are able to crack this oyster. So far I've gotten four (4) other owners to confirm my findings.

Cheers,
-Shark

 

[Hellbinder]

Wow thanks, i really appriciate it.

I have been doing internet searches all afternoon, and can find virtually NO benchmarks/screenshots of 4xs. This is very disturbing to me. Everyone is comparing GF4 ti's standard 2x and 4x with a coupple doing QC, straight to Ati's smoothvision. this is very questionable.

With SV clearly delivering superior IQ, and 4xs for Nvidia delivering superior IQ these should be "THE" methods used for comparring the two. I read (on one of the only sites i can find any real info at all on the subject) than Nvidia's 4xs looks equal to 6X SV. that is a major accomplishment if its true, especially if the performance is better.

very odd, when i was only able to read one or two vague references to 4xs having a bigger performance hit than Nvidias standard, no comparrisons to SV to speak of, and no direct benchmark results to compare.

It makes me wonder what the hell is going on, in review land.

 

[LLB]

Sharkfood,

On the NFSU images there is a subtle difference.
If you follow the road into the distance you will see two "posts" and between them what looks like a blob with 4x but with 8x it's clearly three distinct objects.
I would be interested to see that same screenshot on an 8500 with max AF just to see what detail is missing from the GF shot (assuming there is a noticeable difference between the two).

LLB

 

[Typedef Enum]

Hellbinder,

Go to our site, and check out my Leadtek article. I know...you're probably thiking, "shameless plug." But I took 4xs screenshots of, I think, about 14 different games.

http://www.nvnews.net/reviews/Leadtek_ti4400/page_1.shtml

 

[LLB]

Here's a couple screenshots of Morrowind, to get the full impact though you really need to see it in motion.
The edge crawling is really bad without AA, with 4xS it's virtually eliminated.
I will gladly live with the performance hit with this game, it just makes that big of a difference. (frame rate in lower right corner)
For some reason I could not capture 2xAA screenshots but 4xS is noticeably superior to it.

LLB

http://webpages.charter.net/madman11/MW-4xS.jpg
http://webpages.charter.net/madman11/MW-NoAA.jpg

 

[SvP]

Probably off topic, but 4xS does something with the water. I'd say that the waves are bigger but what's the cause of that?

 

[KimB]

Some clarifications:

Quincunx does not appear to blur textures in the least on my GeForce4. It most certainly blurred them on the GeForce3, however. I tried to take some screenshots, and, unfortunately, I'm unsure as to whether or not I was able to capture the Quincunx shots. Regardless, I don't use Quincunx FSAA as it has problems with text in some games. I actually use 2x FSAA on nearly all my games.

Anisotropic filtering is just screwed right now on the GF4 + Direct3D and I've posted this several times. It was better in previous drivers on the GF3. Right now it does very, very little... does absolutely NOTHING between 4x and 8x in Direct3D, and carries a big hit.

Here are some screenshots I took:

http://169.237.254.233/Aniso4x.jpg
http://169.237.254.233/Aniso8x.jpg

Unfortunately, it's not easy to see the difference in these shots, but just open the shots in two different windows and swap between the two. The difference should be apparent in textures that are further away. I used RivaTuner to switch between 4x and 8x anisotropic.

Btw, you should notice a similarly small image quality difference on a Radeon when switching between 8-degree and 16-degree anisotropic, and there will be many scenes that will show no difference at all (Actually, there will be more scenes that show no difference in this case than in the 4-8 case with a GF4...).

One thing that you have to realize about anisotropic filtering is that it only increases texture quality on the textures that need it. This is the sole reason why anisotropic filtering generally has much less of a performance hit than SSAA. The truth is that some scenes just don't need any higher than 4-degree anisotropic, while others would benefit from 32-degree anisotropic. In general, the more steeply-angled surfaces you have, the more anisotropic you need. The best example is when you are on a perfectly flat surface looking straight across a very long distance.

 

[KimB]

Probably off topic, but 4xS does something with the water. I'd say that the waves are bigger but what's the cause of that?

Different timing of the shot...

 

[Reverend]

I have been doing internet searches all afternoon, and can find virtually NO benchmarks/screenshots of 4xs. This is very disturbing to me. Everyone is comparing GF4 ti's standard 2x and 4x with a coupple doing QC, straight to Ati's smoothvision. this is very questionable.

There are a number of websites who benchmark using 4xS as well as providing screenshots. NVNews has screenshots, and if you've bothered to read a a review at this site, here as well.

As for comparing NV's 2x, 4x to ATI's SV, well, NV's 4xS isn't exactly implemented by ATI. It isn't questionable but rather what you are lookign for in terms of comparisons.

 

[Hellbinder]

Type, Actually yours is one of the "few" i was talking about. Once I got that java applet, i see what you mean. 4xs looks really really nice. I dont even see how 3dfx types could bag on that. There is also clearly no comparrison between 4x with aniso and 4xs. The ansio solution is far far to blurry in comparison imo. It even looks "blurier" than the strait 4x mode.

Why does it not work in openGL?

 

[Hellbinder]

NV's 4xS isn't exactly implemented by ATI

Really? you think?

please...

The point is SV, and 4XS are both designed to achieve similar results. And both have greater impact on performance. Thus, they should be directly compared when possible, on reviews that compare the two cards/chips.

 

[Hellbinder]

Reverend,

By the way nice review.

yours is by far the most informative, with types a close second. Seriously though, Most other reviews do not have a great ammount of detail on 4xs.

 

[Sharkfood]

Chalnoth-

Quincunx does not appear to blur textures in the least on my GeForce4

You might want to consider upgrading your monitor. Viewsonics and MAGS dont show quincunx blurring on a GF3 at all either. Try a high-end Sony or NEC and it's plain as day.

About the only BIG difference I've noticed with Quincunx on my Ti4600 versus my GF3 is that it appears to implemented by post-filter rather than sw/drivers... aside from being *slightly* better. Reason being- the contents of the framebuffer do not match the onscreen display.

You said it yourself- text is blurred and there is no denying that. I'd also argue so are textures (as most other GF4 owners will also testify), albeit slightly less than the GF3. Here are two shots: one with quincunx, one without:
No Quincunx:

Quincunx:

The text in-game is substantially more blurry, as well as the rest of the scene. It looks like they may have found a post-filter hardware solution that is not only a tad more superior, but doesnt have the effect in the framebuffer.

Here are some screenshots I took:

Neither link appears to be working at current. Can ya' create a freebie tripod or something? Only costs you an email address for spam (just make a hotmail to go with it. lol) and ya get 20MB for hiking images.

Btw, you should notice a similarly small image quality difference on a Radeon when switching between 8-degree and 16-degree anisotropic, and there will be many scenes that will show no difference at all

You may actually want to try a Radeon before making false statements like this. The difference between each degree of anisotropic filtering is night and day at each level.

http://shark_food.tripod.com/drugs/drugs.html

One thing that you have to realize about anisotropic filtering is that it only increases texture quality on the textures that need it

So, you are now describing a psychic-method of operation anisotropic filtering? When did the NV-swami chipset come out? lol.

The truth is that some scenes just don't need any higher than 4-degree anisotropic, while others would benefit from 32-degree anisotropic.

Some people would equally argue that no anisotropic filtering is needed whatsoever. It doesnt make the case any more or less valid that 4x->8x is broken on the GF4 and that other levels of anisotropic filtering are doing little to nothing with current drivers. Just because User A with crummy monitor B doesn't think higher degree anisotropic filtering is needed doesn't mean that User B with higher quality monitor C is wrong in pointing out the higher level is broken.

I'd personally prefer to see the bugs fixed rather than make excuses or try and circumvent the problem by declaring "Oh, don't need it anyways!" kinds of excuses. It's almost as equally broken on my GF3 with the same drivers whereas was function great in older drivers. Driver bugs dont need excuses, they need fixes... simple as that.

 

[Sharkfood]

LLB-

On the NFSU images there is a subtle difference.

Sorry, that is a slight change in camera angle. NFS recorded replays dont have a *perfect* camera and it can be off by a pixel.

You notice the edge aliasing/stairsteps on wooden bars mirror that change. That's polygon data, not texture. The road texture (as well as all other textures) are indeed identical.

I noticed on Typedef's review his Shot 5 and 6 mirror the same behavior from the 4xS shots. I think it's pretty visible that 8x in Direct3D is making no visual change at all.

It's a silly argument really- especially to anyone that has owned a GF3. 8x anisotropic filtering DOES make a visual difference guys! Go fire one up and see. I'm a bit overwhelmed that there are a few people willing to stretch 8x AF + D3D to make a case that this is somehow "correct" when it's just so painfully obvious it is NOT correct. hehe.

 

[Ichneumon]

I'd personally prefer to see the bugs fixed rather than make excuses or try and circumvent the problem by declaring "Oh, don't need it anyways!" kinds of excuses. It's almost as equally broken on my GF3 with the same drivers whereas was function great in older drivers. Driver bugs dont need excuses, they need fixes... simple as that.

butbutbut... Nvidia doesn't have driver bugs! Their drivers are perfect I tell j00!

 

[Typedef Enum]

Why does it not work in openGL?

Probably the same reason why FSAA didn't work with the original GTS drivers...It simply hadn't been implemented. There must be some good explanation as to why it couldn't work right out of the box...I'm sure it wasn't quite as easy as one might imagine.

What I'm really _hoping_ for is the next round of drivers include this feature. nVidia stated in an interview some 2-3 months ago that they were working on this, and it would be added sometime down the road.

 

[Dave Baumann]

My theory of why it doesn't work in OpenGl is because I think they are forcing an OpenGL multisample buffer when its being used in OpenGl, and I'm not sure the multisample buffer handles SuperSampling and as 4xS is a hybrid of Supersampling and Mulitsampling it doesn't operate.

Remembering back from some of the ARB notes there was some discussion about changes to the Multisample buffer extension so that it could handle the distinction between supersampling and multisampling and with hindsight I believe 4xS to be the reason why it was discussed.

 

[KimB]

So, you are now describing a psychic-method of operation anisotropic filtering? When did the NV-swami chipset come out? lol.

Um, no, this is the way anisotropic works. Here's a link to how it's implemented in OpenGL:
http://oss.sgi.com/projects/ogl-sample/registry/EXT/texture_filter_anisotropic.txt

If you'll note, the degree of anisotropic is chosen on a per-pixel basis by the hardware. This is just the way anisotropic works.

For example, if you look at a shot where you see a difference between 8-degree and 16-degree anisotropic, you'll notice that a large part of the image shows absolutely no difference. This will be true no matter what video card you are using, or what method of anisotropic filtering.

Some people would equally argue that no anisotropic filtering is needed whatsoever. It doesnt make the case any more or less valid that 4x->8x is broken on the GF4 and that other levels of anisotropic filtering are doing little to nothing with current drivers.

This has nothing to do with subjective viewing of the images. Some images simply do not need higher degrees of anisotropic filtering. This is, for example, one of the problems with nVidia's anisotropic filtering method (i.e. the reason the speed hit is so substantial). For example, in 3DMark2k1's fillrate test, the polys are all flat on the screen, requiring absolutely no anisotropic. And yet, the performance decreases on GF3/4's, indicating that they (incorrectly) apply anisotropic filtering to the scene.

The truth is that every 3D game will show a need for anisotropic filtering, of any degree, in some situation or another. Obviously, of course, each game will be different in how obvious the higher degrees of anisotropic are. What you have to be careful about, therefore, when choosing scenes to showcase anisotropic is that the particular scene you are looking at will show benefits from increased degrees of anisotropic. The best scenes for this are high-contrast textures viewed on a flat surface at extreme distances.

And 8x aniso is *not* broken in D3D. Your own screenshots have shown this. Why do you still refuse to believe it?

 

[SvP]

And 8x aniso is *not* broken in D3D. Your own screenshots have shown this. Why do you still refuse to believe it?

Well, he provided screenshots from Need For Speed and Morrowind.

- in NfS there were some differencies between 4x and 8x, but as he explained these were due to change in camera position.

- in Morrowind there were no differencies between 4x and 8x. If there are any - feel free to point them.

Also if you have some screenshots to prove your point (that there is difference between 4x and 8x in DX), try to link them. Broken links are not the proof