how does RAM work on a dual-GPU card?

[dagas]

Can someone explain or link to an explanation of how the on-board RAM is used on a dual-GPU card like the 9800GX2 or the 4870X2. As I understand it the second RAM is barely used. If it is a 2x512MB card like the 9800GX2 it is limited by the 512MB of RAM just like any 1x512MB would be. Why can't the other RAM be used as well or why is there not just one big RAM that is shared between the cores? If the second RAM is useless why even have it? Why not have 1GB on the first core and no memory and the other since it can't load textures into that one anyway.

Please explain this =)

 

[knowom]

Current dual-gpu cards work by using load balancing between two gpu cores on a video card. The current designs on these dual-gpu cards don't have shared memory so each gpu core uses the required amount of vram per gpu core to render and load balance two scenes together seamlessly. Shared memory wouldn't actually give you any more ram, but rather better management of how the gpu core's could handle the provided available on board ram.

With a shared memory dual-gpu you could do more things with the same amount of on board ram essentially. For example card with 1GB shared vram and 1 gpu core could be using 256vram for one application while the other core was using 768vram for another application. Now with a 2x512MB even though the card has 1GB worth of total ram on board your restricted to 512MB per core so that same situation wouldn't be possible. Shared memory is more economical and flexible, but more complex to design which is likely the leading contributing factor as to why it isn't yet used in dual gpu cards. I hope that helps answer your question.

 

[xcorat]

Other than the drawbacks the guy before mentioned, one of significant draw back in dual card solutions is that since the memory is not shared, both GPU cores has to has their own data in their respectable memory arrays. This leads to heavy memory data redundancy, since most of the times both GPUs use the same data for calculations. Fir instance, 4870X2, having 2Gb memory can actually use only a little more than 1gb as net memory. But hopefuly they will try to implement a shared memry of somekind that will get rid of this problem.

 

[Nuker]

In SFR mode the load of each frame is divided and placed upon each GPU in a checkerboard kind of pattern, which means that the memory is effectively shared sort of. In AFR mode each GPU has its own memory array.