nVidia going for 16 sample FSAA?

L

LeStoffer

Veteran
I think they're talking about what they need to have to make 'lifelike' images. They're suggesting it will take 5x more "power" to do the full scene antialiasing to meet whatever bar they've put to handle aliasing.
 
RussSchultz said:
I think they're talking about what they need to have to make 'lifelike' images. They're suggesting it will take 5x more "power" to do the full scene antialiasing to meet whatever bar they've put to handle aliasing.
Click to expand...

I agree. I think Kurt probably took this FSAA approach because it was the best he could think of. But then on the other hand: Maybe it wasn't a total coincidence looking a bit into the NV-future (beyond NV40)? ;)
 
That presentation was incredibly foolish. The number they derived at the bottom assumes that all of the neccessary performance increases will have a multiplicative effect on one another. While many do, I doubt that, for example, doubling the the requirements for shadow map rendering will have much impact on the fact the rendering is to be done in stereo, since the same shadow maps could be used for both view sides at the same time.

What's worse is what they don't take into account. Nothing is noted about the needed increases in transformation ability, other than an implied assumption it will scale linearly with rasterization requirements (which might be correct but still is far from obvious). Nothing is mentioned about the need for even a rudimentary global illumination scheme, which to me is more important than 4K x 3K resolution or 500Hz (!) frame rates (but then again this is nVidia...)
 
OT a bit here, I was under the impression CRTs had better contrast than LCDs yet that paper has only 300:1 ratio for a crt. How do they actually measure contrast ratio?
 
LeStoffer said:
RussSchultz said:
I think they're talking about what they need to have to make 'lifelike' images. They're suggesting it will take 5x more "power" to do the full scene antialiasing to meet whatever bar they've put to handle aliasing.
Click to expand...

I agree. I think Kurt probably took this FSAA approach because it was the best he could think of. But then on the other hand: Maybe it wasn't a total coincidence looking a bit into the NV-future (beyond NV40)? ;)
Click to expand...

Hmm...
Well, if they did:
8 samples/pixel
2x2 reconstruction
With 32 different sample positions, preferably programmable...

You'd have 32 samples/pixel, thus +- 78 FLOPS/pixel.
It's still quite a bit, but then again, it's only done ONCE per pixel. And if you accept some more latency, it doesn't have to be THAT fast with triple buffering...

I think allowing something like 16+5x5 in the NV40 is not out of the question. I mean, 16 samples seems like a given, and if you allow a 2x2 reconstruction filter, better to allow it all the way up - makes the professional market drool over your card, too!

Whether they'll go that far this time or we'll have to wait for the NV45/NV50 is the question though...


Uttar
 
16X super-sampling with programmeable sampling positions would look really good... Performance might not suck if you only store 1 sample for pixels which are completely covered by a single primitive.
 
Well I would say GOOD. its about damn time that the number of AA samples gets increased. you need that (combined with 60fps) to begin to achieve "the CGI look". 6x and 8x samples simply are not enough.

16x FSAA, I would think, should be the MINIMUM for games, going forward over the next few years.

16x FSAA at 60fps or 32x FSAA at 30fps would be sweet for NV4x and R4xx.

ok, ok 32x FSAA is not likely to happen for awhile in reality. maybe with NV5x and R5xx ? 16x for NV4x and R4xx though for sure. pleeease!


Maybe with 16x FSAA, Nvidia would be able to finally match the anti-aliased quality of the "Raven demo" that was shown in 2000 to hype up the Xbox! or do you think that prerendered CGI demo used more than 16x FSAA ?
 
psurge said:
16X super-sampling with programmeable sampling positions would look really good... Performance might not suck if you only store 1 sample for pixels which are completely covered by a single primitive.
Click to expand...
That would be no better than multisampling then. You don't know if a primitive will cross the pixel or not. If it does, then you need all of the depth and color data for super sampling/
 
16x FSAA, I would think, should be the MINIMUM for games, going forward over the next few years.
Click to expand...

Gotta admit I'd be happy if ATi just continue to develop what they've got now. Maybe up to 8* with some sort of adaptive sample pattern.
 
Megadrive1988 said:
Well I would say GOOD. its about damn time that the number of AA samples gets increased. you need that (combined with 60fps) to begin to achieve "the CGI look". 6x and 8x samples simply are not enough.

16x FSAA, I would think, should be the MINIMUM for games, going forward over the next few years.
Click to expand...
1600x1200 at 16x AA would require over 256 MBs of RAM. Not very practical.
 
OpenGL guy said:
1600x1200 at 16x AA would require over 256 MBs of RAM. Not very practical.
Click to expand...
..unless youve got a tiler of some sort?

EDIT: BTW, im watching a 720x480 @ 30fps movie on my 19" CRT. Looks real enough to me. What distinguishes it from Quake3 isnt lack of insane frame rate, stereo view, or HDR.
IMO, the priorities in this Brute Force preaching are messed up. Proper global lighting solutions, proper spatial and temporal antialiasing should come first for images on computer screen.

As for fully convincing headset-type VR, i suspect direct neural input methods will get there first anyway :/
 
OpenGL guy said:
Megadrive1988 said:
Well I would say GOOD. its about damn time that the number of AA samples gets increased. you need that (combined with 60fps) to begin to achieve "the CGI look". 6x and 8x samples simply are not enough.

16x FSAA, I would think, should be the MINIMUM for games, going forward over the next few years.
Click to expand...
1600x1200 at 16x AA would require over 256 MBs of RAM. Not very practical.
Click to expand...

1600x1200 32bpp 16xAA (assuming MSAA) and 64MB texture data = 317MB of ram.

This doesn't include vertex data or other types of data.

This also assumes the architecture is extremely efficient.

In real world I would estimate around 350MB-400MB of VRAM would be eaten up.
 
I wouldn't be surprised if 512MB cards come into vogue next year as the bleeding edge. :)

Anyways, I'm more interested in Nvidia cleaning up their driver shenanigans moreso than any other added value of their products. As it stands now, I simply don't trust them or their products. That's a pretty huge fall imo considering where I stood just a year ago.
 
OpenGL guy - you're right... brain's not working too well ATM

What about some kind of adaptive super-sampling :

say we start with a minimum of 1 sample per pixel for every covered pixel in 2x2 stamp.

- based on some criteria (such as pixel stamp coverage, texture anisotropy, ddx/ddy of pixel shader results differing by too much), subdivide into 4 or 16 samples per pixel.

Essentially multi-sample when you can get away with it, otherwise super-sample.
 
i wonder if a (3x3) 9 sample AA might be used in the interim for the high end cards of the near future (r400??). What would the image quality improvements be over the current 4x and 6x methods (provided it used similar sampling patterns)?

Would a high end TBDR be the first card with the capability to introduce a usable real-time gaming 16 sample AA until the r500/nv50 tech arrives, late 04- early 05??
 
psurge said:
What about some kind of adaptive super-sampling :

say we start with a minimum of 1 sample per pixel for every covered pixel in 2x2 stamp.

- based on some criteria (such as pixel stamp coverage, texture anisotropy, ddx/ddy of pixel shader results differing by too much), subdivide into 4 or 16 samples per pixel.

Essentially multi-sample when you can get away with it, otherwise super-sample.
Click to expand...
You'll get weird effects on internal edges if you treat them differently than the rest of the polygon. For example, take a quad drawn as two triangles. If you do extra filtering on the edge where the two triangles meet, then it will look funny.
 
You must log in or register to reply here.
Back
Top