You know,
 
I contacted one of the product managers on the forums and asked if they were working on a 3GB version of the card, their answer was:

there has not been enough testing to see if games utilize over 1.5GB of framebuffer with 3 displays.

 
So I figured I'd make a topic about it.  For those out there who have three displays and are eyeing the Palit Microsystems' 3GB variation of the same, perhaps you could share your experiences with large resolutions spanning three displays.  Notably from the reading I've done, the issue seems to rear its head when using anti-aliasing due to the exorbitant amount of memory required to do so.
 
People with experience in how 3D engines work are welcome.  From my limited understanding of FSAA (Full Screen Anti-aliasing), it functions by rendering the scene at twice the effective horizontal and vertical resolution (i.e. 4 times the original size, giving each pixel four data points to interpolate for the resultant pixel.)  On 3x2560x1600 you get a scaled up resolution of 3x5120x3200 or 15360x3200 per frame.  Doing basic math says that results in 196,608,000 bytes per framebuffer if 4 bytes are used per pixel (4*15360*3200), I'm not sure but then you'd have to have a second buffer to display it to the target resolution 3x2560x1600, which is another 49,152,000 bytes per frame (effectively one frame equals 250MB, 1/6th the effective memory of the card.)
 
Since it wouldn't make sense for the cards to transfer full data sets across the SLI bridge, there's likely media duplication on each card which represent the assets used to render a given frame.  This would mean that each card has its own copy of the textures, someone correct me if this is wrong, I think I read it once, it sounds logical.
 
Here's where it gets confusing for me, if each card handles the frames in a round-robin style (via Alternate Frame rendering), the master card, or the main card which handles much of the work, would it need a buffer for each of the slave cards' frames?  I'm not certain on how the SLI functionality works, so when card 3 finishes its frame and sends it to card 1, does 2 allocate space for it or just straight forward it?
 
The answers to these questions might shed some light as to whether 1.5 GB of memory is sufficient for large format rendering or quickly finds itself with a deficit due to the volume of information just contained within each frame (ignoring the assets needed to render said frame.)
 
From a review by [H]ardOCP, the memory available is an issue, on 5,760x1200, no less; if that resolution has an issue, I can't imagine what a factor larger, 7680x1600 would have (1 & 7/9 times the area.)
 
On a fantasy note, if they made a FTW edition, would it be called the GTX 580 FTW Enthusiast's Vision?  Emphasis on the vision part, since it's about 3D Vision.
 
Let me know what you think, perhaps if there's enough enthusiast's interested, we might just see an EVGA alternative to Palit's card.

Alexander Morou

Here's where it gets confusing for me, if each card handles the frames in a round-robin style (via Alternate Frame rendering), the master card, or the main card which handles much of the work, would it need a buffer for each of the slave cards' frames?  I'm not certain on how the SLI functionality works, so when card 3 finishes its frame and sends it to card 1, does 2 allocate space for it or just straight forward it?

 
Even in AFR mode both cards buffer the frame, so 3x 1.5GB cards in SLI mode = 1.5GB usable frame buffer.

My guess is that this card will not happen and especially on this generation of GPUS, if it is very expensive and/or time consuming to develop. The amount of people that would use it is MICROSCOPIC.

jedi95

Alexander Morou

People with experience in how 3D engines work are welcome.  From my limited understanding of FSAA (Full Screen Anti-aliasing), it functions by rendering the scene at twice the effective horizontal and vertical resolution (i.e. 4 times the original size, giving each pixel four data points to interpolate for the resultant pixel.)  On 3x2560x1600 you get a scaled up resolution of 3x5120x3200 or 15360x3200 per frame.  Doing basic math says that results in 196,608,000 bytes per framebuffer if 4 bytes are used per pixel (4*15360*3200), I'm not sure but then you'd have to have a second buffer to display it to the target resolution 3x2560x1600, which is another 49,152,000 bytes per frame (effectively one frame equals 250MB, 1/6th the effective memory of the card.)

To clarify this, what you are talking about is super sample anti-aliasing, or SSAA. This method is rarely used because of the massive performance hit compared to multisample anti-aliasing or MSAA.

More details about the differences can be found here:
http://www.anandtech.com/show/2841/14

You are so wrong its not funny.

I was agreeing with you, and disputing jedi's post actually.
 
the use of super sampling is extremely common these days, especially in games that use Direct3D 9 engines which make use of deffered shading.
 
Not to fail to mention Transparency Supersampling being used more than Transparency Multisampling, adds even more samples being stored.
 
Even more common is the use of hybrid MSAA+SSAA techniques as there are a number of games (and the opengl api itself) that do not support transparency antialiasing, without modification of the engine itself, and these tend to chew up higher amounts of ram than just pure msaa or ssaa modes.

keep in mind, that we are reaching 1GB framebuffer usage with just 4xMSAA with a 1920x1200 output resolution, and this is not just by using alot of high resolution textures, you also buffer HDR and normal maps, HBAO and SSAO also require additional video memory.
 
Just cause 2 is a good example of this, as 16xCSAA (4x+12x) consumes the entire framebuffer.
 
multiple display surround is easily going to be framebuffer limited since no matter how many 580's you have, its still only 1.5GB all up.

For the high-end enthusiast market (i.e. a sub-portion of us), there is a small market for a 3GB card, I'm sure. The question then becomes, is the GF110 with a 3GB frame buffer an adequate balance of compute strength vs memory throughput. I'm sure folks can utilize the frame buffer, Crysis nearly maxes the 580s frame buffer at 1920x1080. So it's a reality especially at surround resolutions. But the real question at that point could the GF110 keep up with it?
 
I'm certain that when the product developers mentioned there hasn't been sufficient testing, this is what they were probably referring to. Sure, slapping more chips on the card makes up to a point. But what is that point where cost eliminates marketability?
 
Personally, when I had my 480s in SLI and the NVIDIA drivers had supersampling (or whatever it was) enabled for the entire frame it was absolutely gorgeous. I was marveling over the scenes in BF:BC2. Once that was fixed there was, to me, an obvious difference. While fps did shoot up, the quality of the image was still great but not as fabulous as it was. I'd easily consider a card with additional framebuffer, if and only if, the GF110 has the compute strength at a price point that is palatable.

Right,
 
Myself as far as the higher-end enthusiast market goes, I research my purchases first.  I weigh the outcomes and purchase based upon what I conclusions I can make given my limited access to the product(s) beforehand (the only actual place within a few hundred miles that has any electronics for purchase, is Walmart, higher end electronics, I don't even know where to look.) 
 
As far as the compute strength goes: the Palit Microsystems' cards seem to give hints that the cards have extra legroom to spare under the right circumstances.  Sure, you might not be able to go to the highest AA settings, but we're talking about resolutions that are 5+ times that of a single 1920x1200 display (nearly 6 times the size of a 1920x1080 display.)  Each frame at that size is also 1.7777--times the size of 3x1920x1200, and almost, as you'd think, twice what it would be 3x1920x1080.
 
So yeah, even Fermi having issues with this resolution isn't any surprise.  I'm half tempted to wait six+ months for the Kepler architecture, especially since there doesn't seem to be a strong enough indication by EVGA at this point to release a 3GB variant.  The longer they deliberate, the closer Kepler gets to launch, and the less valid the decision to purchase them becomes (my resources are more constrained than some higher end enthusiasts, so waiting doesn't bother me.  A $5,700 dollar upgrade, 3x30" displays and 3 Video cards, would take me ~3 months to procure.)

I guarantee you if a 3gb model was made, the computer animating community would buy up this card. None of us want to spend $800,$1500, or $3000 for a Quadro card to render. $600 for a card with 3gb of ram to get renders done plus when not working, can actually game on the card too.. you would be selling these like hot dogs at a baseball game.

Right,
 
EVGA needs a reason to make the card.  Some person online who says professionals will buy it isn't as telling as people actually buying a similar product that isn't theirs.  Two enthusiasts here have already bought three of their competitors products, while that's only six points of purchase, if the number continues to rise, the reason becomes more and more compelling.  So long as all those interested in the card don't buy Palit's first. :P