NVIDIA G80: Image Quality Analysis

Trilinear optimisations affect quality. But you are right that Nvidia is certainly not the only one using them. I dont mind optimisations so long as they are controllable. The trilinear opts in the case of the G80 have not reached complete redundancy as far as performance goes. Though I sincerely question the validity of leaving texture stage optimisations ((Even if they are much less aggressive)) in the performance modes as I have my doubts of their performance benefits.

In my own cynical nature I believe HQ on the G80 is the filtering level we should have "always" been at. However with the level of crappy optimizations that we have become accustomed too over the last 3 years its not really surprising that some remain left over. Especially when they benefit performance still.

 

Well I've always been very sensitive to texture aliasing and I was one of the ones who could notice a slight 'shimmer' increase between my 9800 Pro (Trilinear) and my X800 XT ('Fast Trilinear/Brilinear' AI=ON) but it was only certain very high-frequency textures that showed up this case (like the sand/gravel/wood textures in Half-Life 2).

On the 6800 GT and 7800 GTX 512 I had to enable 'High Quality' (to disable all filtering optmisations).

The G80's Trilinear filtering optimisation (brilinear) is much better but it still is distracting on the textures types mentioned earlyer.

Thankfully G80's High Quality mode is exactly what I require - yippe at last!

 

Hmmm, can someone please explain why CSAA wouldn't work at the intersection of two polygons as demonstrated in the article? Why would that matter?

 

Because for the 16x CSAA mode that looks like 4x MSAA at intersections there are only 4 depth samples. There are 16 coverage samples though and when objects intersect the coverage samples are not an accurate representation of what true 16x AA would look like.

It's the same reason Matrox's FAA had no antialiasing at intersecting objects.

 

link is a 404 for me, I'd like to reach the article.

 

Thanks!

But I still don't think I get it. I'm not sure why you made the distinction that there are only 4 depth samples at intersections - isn't it always 4 depth samples no matter what?

The B3D article doesn't really say anything at all about CSAA's dependencies and what causes it to fail. (It would also help if the navigation dropdown was populated with page headings instead of just numbers )

I've seen it mentioned that the use of stencil shadows disables CSAA presumably because CSAA information is stored in the stencil buffer if that space is available? But I still don't understand why it doesn't work at polygon intersections

 

Note that I don't know for sure how this works, but I'm fairly confident in my assessment. Yes, there are always 4 depth samples, not just at intersections.

Consider two intersecting triangles, each of which cover the entire pixel. Since they cover the entire pixel the coverage information is the same for all subpixels. The only distinguishing information that is left are the depth samples. Since there are only 4 depth samples the triangles can only intersect in 4 ways. Stencil shadows would always be 4x because only 4 stencil values are stored and there's no way to use coverage information to store additional stencil info.

The coverage information will only enhance the quality when there is a polygon edge.

 

Short question: What's the bottleneck for CSAA performance?

 

I did this a while ago. It's not super accurate and I can't quite figure out the reasoning behind some of the numbers, but here goes anyway: http://www.notforidiots.com/AA-Tests.txt

In the coreclock-limited scenario, you'll see how things scale with the ROPs most likely, and in the memclock-limited scenario, you'll see how memory-bandwidth-intensive the AA mode is for a fullscreen quad.

It's probably worth noting that 'bandwidth' there is really only related to the amount of raw data and the maximum compression ratio. As you can see in these numbers, the G80 does not increase its maximum compression ratio with AA - unlike ATI hardware which, AFAIK, does. But on the other hand, the average ratio achieved in realworld conditions will improve, because coherence will be higher.

On a related note, hopefully me and Rys(?) will have the time to do some further digging, including DX10(?), in time for the G84/G86 launch. We'll see how that goes, but no promise!

 

I think that if they added csaa then they should have added the 3dfx voodoo 5's method, for those who want the best iq, at least for sli.

I realise most people wouldn't use it b/c they prefer higher framerates, but those who want the best iq are suffering.

I could be wrong about this, but I think due to the voodoo5 method's ss nature it would also be 100% compatible and hassle-free.

I know that the v5 couldn't do shader fx, but I think that its method would purge shader aliasing in games made in the past several years if it were used in today's hw.

There's a lot of games that msaa doesn't work with.

Also, another thing that's disappointing about the 8 series new hq af, is that the center of the circle still isn't quite 100% round. I think the fact that it's "near" angle independent maybe showing in those af iq tests.

Then, the other thing that kind of sucks, is that there's no way in the driver to globally disable texture compression.

For 95% of games already out, and every open gl game out, it doesn't need it.

I have dual 8800 gtxs in sli, and the image quality could definately be better. Frame rates aren't a concern at all. It's not necessary to everyone to have their frame rates go over 40, nevertheless 60 fps.

 

For the love of God....what is so magical about 3dfx, that a billion years after it`s death, ppl still think that they had the magic sauce, that their FSAA is unparalleled, that the Rampaga would've owned even the G80 with some minor tweaking etc. I can see it, some 300 years into the future, on the deck of a starship similar to Star Trek ones, the pinnacle of human technology, running through the stars, and the captain goes:yup, the screen is pretty good...if only they would've added some 3dfx FSAA to the image, it really would've beaten our super mega giga ultra methods of rendering, developed for years upon years by the brightest minds of the human race. Because 3dfx was simply cool like that.

 

haha, it's just sentiment I think

 

Thanks for not attacking me directly, but there could be some things that a lot of people don't realise. I'm not saying you didn't already know, and everyone else in the world probably disagrees with me about 3dfx having unrivaled aa.

however, the things that should be pointed out is:
1. CSAA/MSAA could still be an option for those who prefer a few extra fps.
2. you're already getting the performance hit for trss w/ csaa and msaa.
3. it simply does edging in lines better with less samples
4. it would purge shader aliasing, which may be purged in future games, but the games out now would still have shader aliasing.
5. it would make shadows a little bit smoother, in some cases, a great deal smoother.
7. if i'm not mistaken, due to 3dfx's methods's ss nature, it would work with all games. csaa and msaa don't work with a fair number of games.
8. the bluriness isn't a problem, just have totally unoptimised tri-af and an LOD mip bias slider.

 

3. It doesn't. For geometry edges, multisampling and supersampling produce equivalent quality.
6. is missing
7. You're mistaken. Supersampling wouldn't work for exactly the same reasons MSAA doesn't work.

 

7. you're probably right, but i could've sworn i remembered playing pop sands of time on my geforce fx with supersampling. and you're absolutely sure for sure that rgss wouldn't work with deferred shading?

3. i would say that if the ss is rotated grid then it does, but if it's ordered grid ss then rgms is just as good for edging. i remember that the geforce fx used og for msaa and it barely made a difference. rotated grid makes all of the difference.

 

3. Rotated grid sampling patterns have been around for ages, ever since the Radeon9700.

But anyway, I`ve learned that arguing with 3DFx fans is dangerous, so I`ll back down.

 

no, that's ok. you know more than i do. i'm sure of it.

anways, if i'm not mistaken the 9700 pro was the first to use rg for msaa, the voodoo5 first used rg w/ ssaa before ati used it with msaa.

also, xmas try to let me know when you have the time.

i'm a dumbass, so that's why i'm so inquisitive.

 

Don`t get me wrong, for it`s time, what 3DFx did with AA was great. It`s simply decrepit now, and unfeasible in the current context. Consider the fact that with SSAA you increase pixel work by your SS factor(be it X2, X4, X8 whatever)-in the context of more and more work being done per each pixel, do we really want that really expensive SSS-mimicking a few hundred instructions long shader to be done 4 times simply in order to get some free AF and to get arbitrary handling of shader aliasing?I`d say no. Also, with SSAA, all of the advanced compression techniques go down the drain, as there's really not much to compress. Whilst MSAA certainly has limitations, and I`d like to see AA techniques evolving beyond the obligatory increase in sample densities, going backwards is not my idea of progress.

 

I could be wrong, but i thought the stream processors didn't suffer any overhead from rgss or csaa or rgms. i thought the rop's do all of the processing for any aa mode.

if the stream processors start doing the aa, then that will take up the stream processors resources.

so either way aa will never be free.

i would just add a shitload of rops, and then you could definately do it, especially in sli.

plus, the frame buffer compression techniques don't yield much real world performance gain. i wish they'd just remove the transistors for all types of compression, b/c that would free up a lot of space. compression sucks.