NVIDIA: Beyond G80...

Extreme ROPs?




Sorry, I just couldn't help myself..

 

Is that accurate? I thought the number of transparency samples is tied to the number of color samples being taken?

 

Thats accurate for SSAA at least trinib. His point was that using 4x transparency SS in a scene loaded with alpha textures may not doing any better than 2x2 Supersampling.

 

I get that part but I was asking about the coverage sampling statement - "The unfortunate thing about Coverage sampling is that you get as much Transparency samples as Z/stencil samples are being stored."

Sorry if I'm missing something obvious.

 

Well the stored color/z-samples are related.. so maybe I am missing something. But if CSAA is 16x and transparency SS is on. It only takes 4 transparency SS. If CSAA is 8xQ or 16xQ, it takes 8 transparency super samples. Since it bases its sample value on stored color/z.

Now I am beginning to wonder what exactly is being discussed myself. Help me

 

Heh, I read Ail's post as "the unfortunate thing about Coverage sampling is that you get as much Transparency (color) samples as Z/stencil samples are being stored."

So that for 16xCSAA you would get 16 Transparency color samples taken for alphas

 

Ahh too funny. I read it as exactly the the opposite. Guess he'll need to clarify!

 

Unfortunately for both of you, each one got half of what I wanted to say.

a) If you have a scene that has a high persentage of alpha test textures, with transparency supersampling you gain only very little in performance as opposed to full screen supersampling (always at same sample densities for both cases).

b) Transparency Multisampling or alpha to coverage (pick your poison heh) on the other hand might not deliver an as high IQ as TSAA/quality AAA, but it's nearly for "free" in terms of performance and a higher than 4x sample density could help closing the IQ gap to 4xTSAA.

c) It is true that with Coverage mask sampling modes you get as many Transparency AA samples as Z/stencil samples are stored. In detail:

8x->4z/stencil = 4xTMAA or 4xTSAA
16x->4z/stencil = 4xTMAA or 4xTSAA
16xQ->8z/stencil = 8xTMAA or 8xTSAA

Now to re-summarize my former crap, I was wondering if with coverage mask modes it would be possible to use as many Transparency Multisampling samples as samples on polygon edges/intersections are being applied (ie for 16x coverage = 16xTMAA). It was actually a question since I can't figure out if it's even possible, but if yes it's a shame not to take advantage of the opportunity.

 

Did they roll over and die with the NV3x disaster? Before someone says no, the "why" would be the key answer to such an unlikely hypothetical scenario.

If sterile GFLOP numbers would only define performance at all times. Here I'd have to wonder why the R580 doesn't slaughter the G80 since their GFLOP throughputs are too damn close. Or do GFLOPs smell differently in D3D10?

 

Isn't 16x -> 4 color + 16 z/stencil !?

 

http://www.rage3d.com/board/showthread.php?t=33872731

http://www.anandtech.com/video/showdoc.aspx?i=2870&p=9

You're sure you're not confusing z with coverage samples?

 

Yeah, I'm confusing something Think it all started with the fact that when stencil data is required at coverage sample positions the number of coverage samples is reduced. So when I saw your statement a bunch of stuff got muddled. But anyway, you were talking about transparency MSAA which I missed to begin with !

 

Well CSAA doesnt work on stencil objects at all. Since only 4 color/Z Samples are taken. Thats why under the circumstances of stencil shadowing. 16xQ can provide quality improvement to stencil shadows.

 

Umm.. who cares? cause this thread is way off topic.

I love how the thread started as someone asking about a midterm upgrade and has progressed to people talking out their a$$ about AA. As for the OP, and the discussion about future nVidia products I just wanted to point out some things that people are aparently missing. To start the G80's (8800GTX/GTS) already support through GDDR4. secondly people are missing the boat on this whole DX10.1 crap. M$ is working on a whole new architure beyond DX but they fubard it like they did with WinFS, but both suposedly will be updates to Vista in SP1 or SP2. As for the R600, right now I don't think any aspect looks promising for AMD/ATI. With the new technologies coming out of intel that they have long term pat.'s on, and nVidia's completly new architure coupled with CUDA, AMD/ATI has a lot of work ahead. personally I think AMD is going to use ATI primarly to provide better in-house chipsets rather than focusing on high-end GPU's. but WHO knows. ok I just gave all the ******s about 6 tangents they can run with... let's see where the thread goes from here.

 

Ah, New Meat with an attitude. I predict a short and fiery tenure if you don't dial the latter back about a notch and a half.

 

Meat did u say?

Looks like meats back on the menu boys! (quote from lord of the rings)

 

I don't know if the thread will hold up for much longer, but in your place I'd be rather worried how long you'll survive here with that kind of attitude

 

I think he has some valid points.. *shrug*

 

I'd have to agree with you on that point, AndrewM. But after being here a while you get used to "all the tangents" that crop up in threads -- goes to show even moderators are 'human'. Nothing wrong with that!

Pharma

 

He might have, had he not posted that shit 15 days after the thread wad dead, and coupled it with rather... unstructured speculation of his own. PopinFRESH, consider yourself warned; if your second post is not less retarded than your first one, you're in for a lot of trouble, because that kind of posting style isn't welcomed here, no matter if some of your points are partially valid.


Uttar