GF3 bench, Quake 3

[Velcrospud]

Tom's hardware:
Quake 3
1024x768x32
159.3 fps

4x AA
51.4 fps

 

[PliotronX]

Damn.. so close to 60fps mark, yet so far =\ I wonder how freaking beautiful it looks though

 

[Dark4ng3l]

it will probably get a bit better with newer drivers(detonator 4??)

 

[Ahriman6]

Tom also probably didn't have anisotropic filtering enabled, which you definitely need since multi-sampling, unlike super-sampling, doesn't address texture filtering. 2x2 AA, Quincunx, and 32-tap anisotropic are going to bring a GF3 to its knees in a game like Quake3 (should be usable for most games, though).

 

[thorin]

"2x2 AA, Quincunx, and 32-tap anisotropic are going to bring a GF3 to its knees in a game like Quake3"

WTF are you talking about, have you read any of the news? What do you think Quincunx is? It's nVidia's new antialiasing system. So how are you going to manage 2x2 AA and Quincunx at the same time?

Also you must have missed their quote:
"...the chip can run Quake III at 1024x768 in 32-bit colour fully anti-aliased at 71 frames per second; at the same settings, the GeForce 2 Ultra could do 34fps. "

They're already at ~52fps, so they need another 20 or so to make 71, with their production/post-production driver/hardware tweaks in the past I'm pretty sure they can make 71.

Sorry if I sounded harsh above, but some ppl here tend to make these great sweeping comments without even knowing what their talking about.

Thorin

 

[Ahriman6]

No, Quincunx is not Nvidia's "new" AA method. The GF3 uses OGMS (ordered grid multi-sampling), with a blur filter (called Quincunx) available as an additional option (that also hits performance by another 20% above and beyond whatever frames are lost by enabling AA). I'm also curious as to whether or not those cited scores are with anisotropic filtering enabled, because multi-sampling does not address texture aliasing, and as such it's not a FSAA solution. It's an edge AA solution. Without high-tap aniso, you're going to have a screen full of texture shimmering and pixel popping (unless the Quincunx blur filter really helps with this).

The problem with these previews from Anand and Tom is that they're mindlessly using Nvidia's PR terms (HRAA, Quincunx), which obviously is already causing confusion to readers. Stop and think why you're not seeing the GF3's AA method being described as exactly what it is, ordered grid multi-sampling! Think about why Nvidia is inventing terms like HRAA for non-technical reviewers to buy into and bandy about without rhyme or reason.

 

[thorin]

Here I'll quote for you from nVidia's own press release:

"The first high-resolution antialiasing (HRAA) GPU, featuring NVIDIA's patented Quincunx AA mode, for high-resolution, high-quality, high-performance multisampling capabilities."

"NVIDIA's new Quincunx antialiasing (AA) mode, generates high performance samples at nearly four-times the rate of the GeForce2 Ultra, while excellent visual quality."

Ok this can lead me to believe that they are still doing OGMS and filtering it since they do repeatedly say "sample" etc in their own press release. However I have also seen figures from them (above) and Voodoo Extreme stating that they do this 3.2 to 4 times faster then the GF2U meaning that it's basically lossless (in comparison). (Recall enabling 2x2, or 4 sample OGMS AA is currently a 75% drop in frame rate, therefore doing it 3.2 to 4 times faster removes this drop *). You also need to note that they always call it Quincunx AA so when they are talking 3.2 to 4 times faster that includes OGMS AA and Quincunx filtering (assuming like you say and the quotes above may lead us to believe/understand they are seperate things).

*
Assume:
GF2U No AA = 100 fps
GF2U 2x2 or 4 Sample AA = 25 fps
GF3 Quincunx @ 3.2 x GF2U Speed = 80 fps

Thorin

 

[Ahriman6]

GF2U No AA = 100 fps
GF2U 2x2 or 4 Sample AA = 25 fps
GF3 Quincunx @ 3.2 x GF2U Speed = 80 fps

Actually, the Quincunx # that Nvidia is currently using is 71fps at 1024x768x32 in Q3. Is that 2x AA or 2x2? We know the Ultra's score was 2x2, probably with trilinear. Note also that super-sampling addresses texture aliasing, whereas multi-sampling does not. So the GF3 will need to have 32-tap anisotropic filtering enabled (possibly only 16-tap if the Quincunx helps with texture aliasing) before scores can be compared fairly.

So before we go comparing scores, we need to know how many samples the GF3 was using and whether or not anisotropic filtering was enabled.

 

[jpprod]

Quincunx filtered 2-sample AA is actually a brilliant idea. Theoretically the quality is very good, and the technique should work very well in reducing jaggies. In fact, with 2-sample Quincunx the card can achieve five degrees of blending on an polygon edge whereas with traditional (OGSS or RGSS) four-sample FSAA there are only four degrees of blending. Also the subsamples are evenly distributed over the entire image, so there shouln't be any artifacts in certain angles. The disadvantage is obvious - neighbouring pixels can affect to each other's colour, and thus there are situations in which image will be blended unnecessarily.

About Cuincunx filter with 4-sample (2x2) FSAA and NV20, there's a thread at Beyond3D BBS which indicates it is an option. Check the Q3 screenshot out, the quality is amazing.

Quincunx FSAA thread at Beyond3D

 

[Ahriman6]

Yes, the quality is good. I wouldn't go so far as to say amazing. It's probably between 2x and 4x RGSS, though closer to 4x. Should we really jump up and down that Nvidia is finally offering AA that's comparable to something that's going on a year old? Hmmm, not sure about that.

I'm not trying to troll or bash on GF3 here, but people need to realize that the bandwidth hit for 2x2 OGMS (GF3) is the same as 2x2 OGSS (GTS), with another 20% performance hit for Quincunx operations (at least with current drivers; this will, hopefully, improve with updated drivers). Add to this the fact that multi-sampling does not address texture aliasing and therefore requires 32-tap anisotropic filtering to really clean up the texture shimmering and I don't think we're looking at a package that's going to offer drastically better or faster AA than we've already seen.

Of course what we will see in the near future are reviewers comparing GF3 edge AA (OGMS) scores against V5/GTS/Radeon FSAA (RGSS/OGSS) scores, which isn't apples to apples (not unless anisotropic is enabled during testing on the GF3).

 

[Chad]

Is it just me or does the quinkunx (or whatever) look even blurrier than both 2x and 4x normal FSAA?

 

[thorin]

"but people need to realize that the bandwidth hit for 2x2 OGMS (GF3) is the same as 2x2 OGSS (GTS), with another 20% performance hit for Quincunx operations (at least with current drivers; this will, hopefully, improve with updated drivers). "

That may be a logical expectation/hypothesis but based on the information in hand it isn't correct.

Everything posted to date about the GF3 states that it does Quincunx AA (whether that's a new method or more likely 2x2 OGMS with filter) at 3.2 to 4 times the speed of a GF2U.

We know that a GF2U does 2x2 OGMS with a ~75% performance hit. Therefore in a game where a GF2U hits 100fps enabling 2x2 OGMS would cut your fps down to ~25. Thus in the same game with the same settings a GF3 would score 3.2 x 25 = 80fps to 4 x 25 = 100fps.

"Add to this the fact that multi-sampling does not address texture aliasing and therefore requires 32-tap anisotropic filtering to really clean up the texture shimmering "

I agree with this they will have to enable some aniso to clean up texture issues.

Thorin

 

[Ferocious]

quincunx sux.

Looks like 4X-FSAA for FPS's aint gonna happen for awhile yet.

I can't believe the GF3 is only getting 51 fps here. Maybe 60 with more mature drivers later, but the ancient technology V5-6000 was going to do that.

 

[thorin]

"but the ancient technology V5-6000 was going to do that. "

Yup is WAS going to SUPPOSEDLY on PAPER. To bad it was all just marketing BS.

Thorin