Crossfire Super AA vs. SLI AA

radeonic2 said:
Ya looks that way.
openglguy thinks its draw distance/lod, which would make sense.
Click to expand...

draw distance makes sense for the bushes on the ground in the distance, which are being clipped by the source engine. It does not make sense for the bush/tree on the roof of the building to the left of the truck. In that part, the entire object is not being clipped, rather parts of the texture are winking out.
 
DemoCoder said:
draw distance makes sense for the bushes on the ground in the distance, which are being clipped by the source engine. It does not make sense for the bush/tree on the roof of the building to the left of the truck. In that part, the entire object is not being clipped, rather parts of the texture are winking out.
Click to expand...
so you think it's some kind of optimization?
 
what I'd want to see is bench results for an "unoptimized games", where SLI is slower than no SLI :).
I'd expect then that Super AA performance is almost on par with single card performance (but with the benefit of added 2x supersampling)
Same could be said of "optimized games" as well.

(actually if I got suddenly rich I would get two 7800GTX only to run them in Super AA all the time and never other SLI modes ;))
 
I didn't look at the article too closely, but did they mention if they enabled gamma correction on the nVidia drivers?

Nite_Hawk
 
chavvdarrr said:
comparing speed when one makes 2xSSAA and other 4xSSAA ... and wondering why first is faster !?
Click to expand...

yes, and ATI does 12x (6x MSAA + 2x SSAA) where NV does 16x (4x MSAA + 4x SSAA).

I'm not sure about the sampling patterns though. I'm quite sure it's ordered grid for Nvidia but I don't know for ATI (do they do rotated grid with image compositing similar to what happens on voodoo5? or is it two rotated grid images which are "ordered-gridly" combined? or do they revert to all ordered grid like I think nvidia does for 8xS and such)
 
I read the review last night and frankly that is not overly new info and the X8xx crossfire also performs the same versus GF6 SLI . But ............. credit to Techreport and Firingsquad for being the only sites so far willing to do a head to head beyond 4x AA .

Why pay over $850 for a pair of video cards which can do insane ammounts of AA ( 10 to 16 X ) at over 45 fps in several games and play at 4x with 130 fps ............. ???
 
interesting to see SLI AA is useless on geforce 6 but runs with expected performance on geforce 7, didn't know that. Is that because gf6 sends framebuffer stuff through PCIe in this case whereas gf7 does it through SLI cable?
 
malficar said:
I like this commentary:

http://techreport.com/reviews/2005q3/ati-crossfire/index.x?pg=12

SLi AA loses gamma correction that a single 7800 card will do. Personally, I share this reviewer's opinion on the selection of sampling points as well.

Performance numbers from a Hothardware review:

Click to expand...


LOL " This occasion has also given me the opportunity to create what I believe is the most incredibly Byzantine results graph in the history of TR. I couldn't be prouder. Gaze upon it, ye who seek enlightenment, and behold its arcane glory. "
 
Blazkowicz_ said:
yes, and ATI does 12x (6x MSAA + 2x SSAA) where NV does 16x (4x MSAA + 4x SSAA).

I'm not sure about the sampling patterns though. I'm quite sure it's ordered grid for Nvidia but I don't know for ATI (do they do rotated grid with image compositing similar to what happens on voodoo5? or is it two rotated grid images which are "ordered-gridly" combined? or do they revert to all ordered grid like I think nvidia does for 8xS and such)
Click to expand...

Actually NVIDIA is using two identical rotated grid samples for each cards ,the center of samples are offset a bit to make a total of 8x samples. Ati has progameble sample pattern, they are using two different sample patterns for the two cards to make an optimal combined pattern - the center for the patterns for the two cards can either be identical (what Ati calls 8x AA) or offset (what Ati calls 10x AA), the link malficar posted has good info on patterns:

http://techreport.com/reviews/2005q3/ati-crossfire/index.x?pg=12
 
Okay, I've looked at the pixels zoomed in, and here is what I can gather:
  1. NVidia isn't doing any gamma correct blending between samples from one card or the pixels between cards. This is why the powerlines are darker. Both cards affect approximately the same pixels while rendering the power lines.
  2. The above is responsible for the major difference in tree appearance. For NVidia, the perceptible jump from sky to grey is much bigger than the jump from grey to black, especially when zoomed out. The tree branches appear to be alpha tested (only 2 shades), so 2x supersampling is the only source of AA.
  3. ATI has a slightly different threshold for the alpha test (I've seen this in other games as well). This may be due to gamma textures, an optimization (though statistically insignificant IMO), or just a plain bug. This is why their are fewer total pixels affected by the tree branches. This is also causing a difference in tree appearance (though less than the previous point).
  4. ATI gradually fades out bushes near the draw distance threshold, but NVidia doesn't. Since this must be controlled by software, I have no idea why this is the case. However, you can expect bushes to suddenly appear with NVidia.

I would say due to the first, second, and fourth points, NVidia will have more popping and aliasing. Due to the third, ATI will have less detail when alpha testing. The first and second could favour NVidia if you have a crappy gamma curve on your LCD.
 
Gamma correction doesn't alter the amount of popping and aliasing, it just alters the distribution of the color ramp for inbetween shades. The total number of different shades and EER shouldl be the same.

The fourth situation, I analyze differently. *Some* of the bushes look like they have been gradually been faded out at different distances, but some of the others look like they've been culled completely, and frankly, I don't find the draw distance at which these billboard sprites are being faded or dropped to be very compelling.

Is this really an artifact of the ATI drivers? I just can't imagine the drivers doing this except by app-detection and some really wonky hack. It must be the game engine somehow using a different rendering path/setting for ATI.

Also, the idea of that non-2.2 non-trinitron gamma curves are "crappy" is the wrong way to look at it. CRT display technology from the 1980s is no longer the gold standard. These days we have CRT, LCD, DLP, OLED, PDP, and upcoming SED/FED. Trying to shoehorn these displays into (imho INFERIOR) IQ parameters would be wrong. Better, is that future graphic cards be architected to deal with the gamut and gamma curve of the underlying display dynamically, rather than fixed.
 
DemoCoder said:
Better, is that future graphic cards be architected to deal with the gamut and gamma curve of the underlying display dynamically, rather than fixed.
Click to expand...

A round of applause from me ;)
 
Tim said:
Actually NVIDIA is using two identical rotated grid samples for each cards ,the center of samples are offset a bit to make a total of 8x samples. Ati has progameble sample pattern, they are using two different sample patterns for the two cards to make an optimal combined pattern - the center for the patterns for the two cards can either be identical (what Ati calls 8x AA) or offset (what Ati calls 10x AA), the link malficar posted has good info on patterns:

http://techreport.com/reviews/2005q3/ati-crossfire/index.x?pg=12
Click to expand...

True, but the SSAA sample positions seem to close for my taste. I don't know how "usual" 16xS (4xRGMS + 4xOGSS) performs on a SLi setup compared to 16x SLi-AA, but the 8*8 EER the first delivers should be efficient enough. Theoretically 16x SLi-AA would mean 16*16, but with such an awkward sample placement I wonder if plain 16xS wouldn't be a better idea after all.
 
Ailuros, I made this graph a bit back and it should show you how SLIAA and the hybrid modes such as 8xS and 16xS compare relatively to the SLIAA modes. The non SLIAA modes are using AFR2 rendering. Maybe not completely useful. But I figured it might answer one of your questions. The quality between SLI16x and 16xS is so close I cant distinguish them anyway. But I am sure they may exist to some with a more observant eye than my own. Keep in mind Gamma Correction works perfectly using the xS AA modes. Unless I absolutely cannot get 16xS/8xS to scale ((which is rare)) I find the SLIAA modes fairly redundant. But useful in those non scaling scenerios.

*edit* Oh btw this is using my 7800GTX SLI setup with an A64 3800+.



aaperf.png
 
Last edited by a moderator:
ChrisRay said:
Ailuros, I made this graph a bit back and it should show you how SLIAA and the hybrid modes such as 8xS and 16xS compare relatively to the SLIAA modes. The non SLIAA modes are using AFR2 rendering. Maybe not completely useful. But I figured it might answer one of your questions. The quality between SLI16x and 16xS is so close I cant distinguish them anyway. But I am sure they may exist to some with a more observant eye than my own. Keep in mind Gamma Correction works perfectly using the SLIAA modes. Unless I absolutely cannot get 16xS/8xS to scale ((which is rare)) I find the SLIAA modes fairly redundant. But useful in those non scaling scenerios.

*edit* Oh btw this is using my 7800GTX SLI setup with an A64 3800+.



aaperf.png
Click to expand...


Well thanks for verifying my suspicions when it comes to performance at least since IQ can be somewhat subjective. Frankly I have severe doubts that the 4xRGSS part of 16xSLi has better overall results than 4xOGSS in 16xS, since in the first mode the SS sampling patterns (albeit rotated grid) are way too close to each other.

aths from 3DCenter had an excellent article over antialising masks here:

http://www.3dcenter.de/artikel/anti-aliasing-masken/

Don't know if you can box yourself through reading it with an online translator.
 
You must log in or register to reply here.

Similar threads

C
Multi-GPUs AFR vs frame lag & smooth animation
Replies
0
Views
2K
Cubitus
C
Kaotik
World of Warcraft: Cataclysm
10 11 12
Replies
238
Views
47K
Kaotik
Kaotik
Back
Top