GeForce FX Launch Details & Interview

The whole anisotropy definition has become a mess ever since ATI introduced the ridiculously high numbers (128X my ass). The number of samples is the right thing to give, it at least gives you an indication of what kind of quality you can expect.

 

Well, when Geoff talks about the "aggressive" settings in the slider, this doesn't sound much different from what ATI's doing, although ATI takes a maximum of 16 samples to nVidia's 8. Presumably AF quality will be covered here as well. The only reason I can think that nVidia might limit it to 8 samples is performance. Ditto, FSAA performance--with regard to FSAA and the implied bandwidth benefits (which nVidia seems rather mute about) of so-called lossless compression, I'm surprised that 8x OGSS is the best nVidia can squeak out--again probably for performance reasons.

 

Ehmm more than 8 samples per pixel would be slightly problamatic given memory constraints too.

 

Well its all really down to how well you do the adaptive part. The GF4s AFAIK are pretty dumb in terms of applying AF meaning they take lots of samples even though they might end up reading the same data over and over (flat on polygon). The trick is that an extremely low percentage of pixels on the screen actually requires 8 or more samples per pixel, obviously you also need to move up MipLevels when going up with anisotropic filtering since else you end-up oversampling something thats blurry to start with. And obviously you need to make good use of the memory bandwidth you have available.

I am no soo sure about some of the talk I have heard in relation to analysing the texture contents and deciding to downgrade filtering settings...

K-

 

No I meant for the FSAA.

 

...And I think you also meant memory footprint too, not bandwidth, correct?

If we want a reasonable memory pool for textures, More than 8X FSAA (via multisampling on a traditional IMR) is certainly going to require much more memeory than 128 MB, especially at higher resolutions....

 

Ahem, when did ATi introduce "ridiculously high numbers"? ATi has always talked about 16x aniso. If you're going to complain, aim that at nVidia, they introduced the "64tap" thing that doesn't even make a whole lot of sense anyway.

 

Mea culpa, I thought it was ATI which used that kind of numbering but it was only the people making the tweaker tools ... presumably as a response to NVIDIA. My memory deceived me.

 

I thought the whole adaptive issue meant that the number of samples taken will not be uniform. If you mean it should not be degree (of rotation) dependent then i agree with you but my opinion is that ATI clearly stated that the number of samples taken were infact dependent on degree of rotation. The difference between radeon 8500 and r300 's implementation of anistropy is that r300 is able to apply anistropy with trilinear PLUS it can apply anistropy at an angle of 45 (altho the number of samples will be far less than ideal/selected) whereas R8500 can not do anistropy with trilinear and the adaptive algorithm in 8500 takes 0 samples at 45 degrees hence making it really blurry.

The reason for this was given by nvidia actually when they said that their anisotropy was better (inspite of being lower quality 8x vs 16x ) cause they could do that in conjunction with trilinear so in effect according to ATI the lowered anisotropy at 45 degress should not be too obvious as it is complimented by trilinear filtering.

All of the reviews I have read seem to indicate that this is indeed the case the level of anisotropy is indeed a lot less than 16 x but it is not overly noticeable. ( I believe digitlife did a specific review on the anisotropy of R300 with a lot of pictures)

 

Uhhh

wtf is 'colour'?

 

Brit for color.

 

Canadian, too.

 

excellent article!

i found some stuff in there downright hilarious though, like nvidia's trying to dodge some very central issues. huge compliments to the writer for asking the right questions

i especially liked: "We're not disclosing the algorithms that we use, but in the control panel we will have sliders." that cleared the matter right up

one thing that's always bugged me is people just calling a mipmap lod bias "anisotropic filtering". check http://www.sgi.com/software/performer/brew/anisotropic.html for more details on what SHOULD be done. i seriously doubt anyone implements this, and i'd like to know what the official word is.

hiho humus! nice site!

.lycium
http://lycium.cfxweb.net

 

Actually that is not a very nice way to implement anisotropic filtering, not that there is really a right way ... but that one is more wrong than most

 

I tested my GF2 MX and GF2 GTS back in the day with a fillrate program someone posted on the old B3D forums.

It doesn't even apply trilinear until it needs to. Flat (perpendicular, flat to the camera) textures never get more than 8 samples (4 for bilinear, times two mipmaps).

So unless they broke that functionality since then...

 

2. The second reason, which may not have jumped out at you upon seeing the name, is that “FXâ€￾ is taken from “3DFXâ€￾. As I mentioned in the introduction, NVIDIA has been working long and hard on integrating the wealth of knowledge (and patents), which they’ve acquired from 3DFX.

Just a hint.........

 

I can take some responsibility for this 128x thing that spread around the web in the original Radeons days. It turned out you could set the MaxAnisotropic registry setting higher than just 16, you could set it all the way up to 128. Setting it to 128 would give more anisotropic effect, but added more aliasing too. Sometimes it looked better, sometimes it didn't. So I added it to the back than famous piece of software Raid-On Tweaker and then the word of course spread quickly around the web. Loads of people thought that it actually was 128x anisotropic.

 

Re: excellent article!

Thanks!
Yours too

 

Hmmm...the answer given about the sliders regarding your question about Aniso is in the latest 41.03 Quadro drivers. So does this mean the algorithms work in current Geforce4 Ti cards now and how well. Here is a link of a screenshot from THG of the drivers using the same control panel as the 41.03 with sliders on NV30 41.34...


http://www.tomshardware.com/graphic/02q4/021118/geforcefx-05.html

[edited for sentence correction]