CraptacularOne

thinkfly

Rich Z
Well, I guess I'm just confused about all this concerning memory "stacking", or not...

What does SLI actually DO then?  I thought it divided the workload between the two GPUs that were ganged together.  And how does it divide the workload?  Well, I thought it would divide the pixels needing to be pumped to the screen by the number of GPUs now handling the workload.  And where is that pixel info actually stored? Well I THOUGHT in the VRAM for each of the GPUs in SLI configuration.  Is this not the way it works?  And if not, then why bother having separate VRAM for each GPU in those cards (ie: GTX 590) that have two processors on one board if there is no gain to doing so?  Surely the manufacturers wouldn't just waste that additional RAM by putting it on the board, now would they?

Sorry if I am overlooking something obvious.  Back in the day when Voodoo first used SLI, they called it Scan Line Interlacing, whereby each processor took alternate scan lines from the monitor thereby speeding up the display by dividing the workload in half.  Maybe this isn't how it works any longer......

A while ago both NVIDIA and ATI used split-frame-rendering (SFR), which was to divide a screen into parts, reducing the resoultion processed by each GPU. In such case vram did stack up in SLI and CF. However SFR was quickly dropped by both companies, as it brought too much problems, mostly due to complicated bugs and ridiculous cost introduced to driver maintenance.

Now the mainstream way of doing SLI and CF is by alternate-frame-rendering (AFR), which is essentially multiple GPUs rendering the same full screen, but they output the rendered picture in turn. This is why vram does not stack up nowadays. AFR is a lot easier than SFR for those writing drivers, though if the output of each frame is not evenly distributed, it would lead to microstuttering.

VRAM has never stacked or added in SLI or Crossfire. The method of split frame rendering did not do anything to combine the memory of each GPU. Each GPU must have access to the same data if they are expected to be able to render the same scene at the same point in time. So all GPUs in an SLI or CF array are using the same data this give you an effective amount that equal to one card. For instance if you have 2 GTX480's and you're seeing a 1300MB VRAM usage reported, both of the card's memory banks are filled with 1300MB of the same data. As I said this is a requirement of having the cards being able to render the same scene at the same point in time.

The methods of diving the work load don't do anything to change this since both GPUs must still have access to the same data in order to render the same secene. Split frame, alternate frame or tiled rendering have no impact on how VRAM is used.

For this I don't agree with. SFR is an out-dated technology. When it was used, there was no need for each GPU to have a local copy of everything. Yes I agree that they must share some common data, but not necessarily all. That is, vram stacks up a bit (not necessarily 100% scaling for each additional GPU though).

thinkfly

CraptacularOne

thinkfly

Rich Z
Well, I guess I'm just confused about all this concerning memory "stacking", or not...

What does SLI actually DO then?  I thought it divided the workload between the two GPUs that were ganged together.  And how does it divide the workload?  Well, I thought it would divide the pixels needing to be pumped to the screen by the number of GPUs now handling the workload.  And where is that pixel info actually stored? Well I THOUGHT in the VRAM for each of the GPUs in SLI configuration.  Is this not the way it works?  And if not, then why bother having separate VRAM for each GPU in those cards (ie: GTX 590) that have two processors on one board if there is no gain to doing so?  Surely the manufacturers wouldn't just waste that additional RAM by putting it on the board, now would they?

Sorry if I am overlooking something obvious.  Back in the day when Voodoo first used SLI, they called it Scan Line Interlacing, whereby each processor took alternate scan lines from the monitor thereby speeding up the display by dividing the workload in half.  Maybe this isn't how it works any longer......

A while ago both NVIDIA and ATI used split-frame-rendering (SFR), which was to divide a screen into parts, reducing the resoultion processed by each GPU. In such case vram did stack up in SLI and CF. However SFR was quickly dropped by both companies, as it brought too much problems, mostly due to complicated bugs and ridiculous cost introduced to driver maintenance.

Now the mainstream way of doing SLI and CF is by alternate-frame-rendering (AFR), which is essentially multiple GPUs rendering the same full screen, but they output the rendered picture in turn. This is why vram does not stack up nowadays. AFR is a lot easier than SFR for those writing drivers, though if the output of each frame is not evenly distributed, it would lead to microstuttering.

VRAM has never stacked or added in SLI or Crossfire. The method of split frame rendering did not do anything to combine the memory of each GPU. Each GPU must have access to the same data if they are expected to be able to render the same scene at the same point in time. So all GPUs in an SLI or CF array are using the same data this give you an effective amount that equal to one card. For instance if you have 2 GTX480's and you're seeing a 1300MB VRAM usage reported, both of the card's memory banks are filled with 1300MB of the same data. As I said this is a requirement of having the cards being able to render the same scene at the same point in time.

The methods of diving the work load don't do anything to change this since both GPUs must still have access to the same data in order to render the same secene. Split frame, alternate frame or tiled rendering have no impact on how VRAM is used.

For this I don't agree with. SFR is an out-dated technology. When it was used, there was no need for each GPU to have a local copy of everything. Yes I agree that they must share some common data, but not necessarily all. That is, vram stacks up a bit (not necessarily 100% scaling for each additional GPU though).

There is nothing to disagree with me about, it's a fact, not an opinion of mine. Multi GPU has always required each card to carry it's own copy of the data it's working with. Split frame rendering does not share memory resources between cards. It never has and probably never will. The method of dividing up the work load doesn't have any bearing what so ever on how VRAM resources are used. Split frame rendering still requires both cards memory banks to contain the same data since they are all rendering the same scene at the same time. Without both cards memory containing the same data it will be impossible for the GPU to render the same scene. I don't think you understand what you are talking about. Think about it for a second: If both cards are rendering the same scene, they would both need to be reading from the same data. Since each cards memory is located locally on the card it is impossible for a GPU to be able to read another GPU's memory bank in a fast enough way t be of any real benefit. Memory runs much faster than any bandwidth that has ever been available over the PCI-E bus or the SLI bridge. The SLI bridge is only good for approximately 1GB per second of bandwidth and even the fastest PCI-E 3.0 bus is only good for another 32GB per second. That is barely enough bandwidth for a 6800GT from 2004. And that's not even taking into account the additional bandwidth needed for the cards to communicate to be able to render the same scene and in what order. If you take this back to when Nvidia 1st introduced SLI with the 6 series it was impossible because back then we only had PCI-E 1.0 and 8GB per second of bandwidth.
 
In any iteration, SLI or CF have never been able to multiply VRAM with additional cards. It's a limitation of current technology. Until we have bus speeds that can approach the speeds of memory on GPU's it will continue to be an impossibility of multi GPU arrays. This is why SLI works the way it does today and the reason why cards cannot share memory.

Locking per OP request...