Excellent article on NVidia and ATi Optimisations
- Thread starter g__day
- Start date
WaltC,
in this review you can see the difference between Nvidia AF and ATi AF:
http://firingsquad.gamers.com/hardware/imagequalityshootout/
Nvidia's AF is simply superior.
in this review you can see the difference between Nvidia AF and ATi AF:
http://firingsquad.gamers.com/hardware/imagequalityshootout/
Nvidia's AF is simply superior.
Bouncing Zabaglione Bros.
Legend
Exxtreme said:
And yet right underneath, the article states:
"The anisotropic filtering shots are not very conclusive. Both images are of comparable quality ; even blown up to 200 percent, the variance between them is too subtle to note."
Bouncing Zabaglione Bros. said:And yet right underneath, the article states:
"The anisotropic filtering shots are not very conclusive. Both images are of comparable quality ; even blown up to 200 percent, the variance between them is too subtle to note."
That was only for 3D Mark 2003 (Nature scene). Look at the Nascar screenshots and especially the crowd and you'll see a rather large difference between Nvidia and Ati (same problem although not as visible in IL 2 - Sturmovik). Angle problem ?
Please use the term "adatpive" correctly!!
The adatpive terminology is not intended to denote the angle rotation element of R300's filtering, but the fact that it takes only the require number of samples for a surface dependant the angle it is relative to the viewport - i.e. if you are looking at a horizontal floor in an FPS then it will likely be sampling at the maximum number of samples taken, however if you are staring at a way that is running parallel with the view port then there is no need to take up to the maximum number of samples. Thats what is meant by "adaptive", the rotation issue is an implementation specific issue.
WRT to the FS shots - you you actually play Nascar from that position?
The adatpive terminology is not intended to denote the angle rotation element of R300's filtering, but the fact that it takes only the require number of samples for a surface dependant the angle it is relative to the viewport - i.e. if you are looking at a horizontal floor in an FPS then it will likely be sampling at the maximum number of samples taken, however if you are staring at a way that is running parallel with the view port then there is no need to take up to the maximum number of samples. Thats what is meant by "adaptive", the rotation issue is an implementation specific issue.
WRT to the FS shots - you you actually play Nascar from that position?
Bjorn said:DaveBaumann said:
Ok, so if i get this correct, the implementation of the "adaptive" filtering has flaws, thus we get these rotation issues as a sideeffect. Maybe we should call it "adaptive with sideeffect" then
The "adptative" filtering is not inherantly connected to how ATI decided to handle the rotatation issue. Both ATI and NVIDIA's implementations are adaptive, ATI has a separate issue wih Z rotation.
Doomtrooper
Veteran
Filtering at all angles is not a requirement for anisotropic filtering, otherwise it would be called Isotropic filtering (equal at all angles)
Now there is instances where Nvidias implementation maybe superior IQ wise, but it also too slow to be considered useable, and the differences are so slight is it really worth a 56% performance hit
(If it is superior why is Nvidia disabling that 'feature' in UT 2003 )
I might add that no one mentions that Nvidia is still limited to 8X, and THAT is something I find very noticeable between ATI and Nvidia, especially Mafia and Racing games where the extra sampling is very notieceable.
Now there is instances where Nvidias implementation maybe superior IQ wise, but it also too slow to be considered useable, and the differences are so slight is it really worth a 56% performance hit
(If it is superior why is Nvidia disabling that 'feature' in UT 2003 )I might add that no one mentions that Nvidia is still limited to 8X, and THAT is something I find very noticeable between ATI and Nvidia, especially Mafia and Racing games where the extra sampling is very notieceable.
Eh? "Adaptive" is used to describe one element of the filtering properties - both ATI and NVIDIA utilise a similar thing (since this is the only sensible way to do AF). The Z rotation is entirely separate and specific to the implementation choices that ATI chose to make with R300.
Doomtrooper said:Filtering at all angles is not a requirement for anisotropic filtering, otherwise it would be called Isotropic filtering (equal at all angles)
What does requirements have to do with this ?
It's not required of Nvidia to make their FSAA implementation a lot better. But it would still be a good thing if they did don't you think ?
Now there is instances where Nvidias implementation maybe superior IQ wise, but it also too slow to be considered useable, and the differences are so slight is it really worth a 56% performance hit
It's perhaps slower then Ati's AF but i would hardly call it unusable.
You're correct but AFAIK, the rotation issue gets even worse then since the difference between the applied AF with different angles will be even bigger at 16 X.
Let me say this again, imo, it's much more important for Nvidia to fix their crappy FSAA then for Ati to fix this thing. Doesn't mean that i don't think Ati should do something about this "problem" though.
If nvidia's AF is too slow so you can set it down. You have the choice how much you sacrifice image quality to get higher framerates.Doomtrooper said:Now there is instances where Nvidias implementation maybe superior IQ wise, but it also too slow to be considered useable, and the differences are so slight is it really worth a 56% performance hit
With ATi's AF you have no chance to filter certain angles with more than 2x AF.
Hehe, Nvidia's AF produces better results in flight sims and games with "natural" game scenes.Doomtrooper said:
ATI's AF is a compromise and you cannot disable it.
Doomtrooper
Veteran
Bjorn said:
The definition of 'anisotropic' for starters.
FSAA does not have a 'definition' on being 'correct'.
Well Beyond3D has shown as much as 56% performance hits with Quality AF on a FX, and I browse Nvnews quite often and see lots of complaints:
FX 5800 on BF 1942
1024x768 , no AA/AF
2003-08-13 15:49:02 - BF1942
Frames: 184672 - Time: 1054641ms - Avg: 175.104 - Min: 64 - Max: 280
1024x768 , 4xAA/8xAF
2003-08-13 16:08:17 - BF1942
Frames: 46513 - Time: 766422ms - Avg: 60.688 - Min: 30 - Max: 93
1024x768 , 4xAA/ no AF
2003-08-13 17:11:02 - BF1942
Frames: 18857 - Time: 224187ms - Avg: 84.112 - Min: 31 - Max: 101
1024x768 , no AA/ 8x AF
2003-08-13 16:23:35 - BF1942
Frames: 5715 - Time: 50422ms - Avg: 113.343 - Min: 28 - Max: 190
1600x1200 , no AA/AF
2003-08-13 16:42:55 - BF1942
Frames: 54572 - Time: 573469ms - Avg: 95.161 - Min: 29 - Max: 114
1600x1200 , 4xAA/8xAF
2003-08-13 16:56:18 - BF1942
Frames: 1137 - Time: 94344ms - Avg: 12.051 - Min: 3 - Max: 23
1600x1200 , 4xAA/ no AF
2003-08-13 16:58:55 - BF1942
Frames: 2669 - Time: 93406ms - Avg: 28.574 - Min: 22 - Max: 35
1600x1200 , no AA/ 8x AF
2003-08-13 17:04:52 - BF1942
Frames: 14509 - Time: 286860ms - Avg: 50.578 - Min: 21 - Max: 70
At 1600 x 1200 enabling 8X AF makes the game unplayabe at a average of 12 fps
I will be honest, I've never seen the rotation issue in any game I've played.
Agreed..always room to improve.
Doomtrooper
Veteran
Exxtreme said:
Lowering the slider is certainly a way of life with Nvidias implementation, in most games a 56% penalty is too much, and that is only 8X.
I compared Il-sturmovik on a 5600 vs my 9700, never saw the 'superior' image quality you mention.
I'm not saying ATI should not improve their method, I also feel Nvidias method is not correct either.
The trade-off to me, especially racing games is worth the slight z-rotation issue.
What does the definition of "anisotropic" tells us then ?
Not that i know much about it but i'm guessing that you could do a lot of things within the definition that would'nt exactly be appreciated by the consumers.
Exactly, so why bother about the definition when it's the end result that counts ?
Isn't the performance hit a "little bit" less relevant then the actual framerates when it comes to determine playability ?
Ok, so finding one game that's unplayable at f.e 1600*1200, 6X FSAA on a R9800 makes the FSAA unusable on that card ?
Yes, but when you're using a geforce, you have allways the choice between high-quality-image and low performance and low image quality and high performance... except UT2003.Doomtrooper said:
I find, to give the choice in the hand of the customer is the correct way to handle this.
Bjorn said:
This proves only two things to me--it depends on the game as to whether or not you can see any difference between ATi's 8xAF and nVidia's 8x AF. I doubt the scene and camera angle used in 3dMK03 are the only ones at which no difference might be apparent.
The other problem here is that ATi offers a 16x AF, which was not compared in the FiringSquad article. It would be fair to directly compare them for purposes of IQ since 8x AF is nVidia's maximum and 16x AF is ATi's maximum AF setting. Just on that basis alone you cannot use this article to infer that "nVidia has better AF" since that was never demonstrated in the article--only that in some cases nVidia's 8x AF was better than ATi's mid-level AF setting of 8x--but nothing about the AF quality between the products was demonstrated such that a general conclusion might be made.
