Why exactly isn't VRAM combined with two cards?

[Jarrod]

Always wondered why 3GB+3GB does not equal 6GB. I'm sure there is a very valid reason, but I always thought. Does that mean that 3GB is going empty, or maybe 1.5GB on each card is going empty.

Just wondering .

 

[Deceptor]

I haven't looked into it but I'd presume because it'd be too slow/tricky to find where data is stored. Plus it could get even more complex with data being distributed to the two cards (eg a texture is stored 30% on 1 card and 70% on another).

This would mean the cards would eat bandwidth just talking to each other so GPU1 can read GPU2 memory. It makes sense in that respect to just mirror the memory and make SLI/CF a much easier system.

Again however I've not looked into it so I could be completely wrong (but as a dev this makes sense to me).

 

[OptimaLnrg]

If the data needed for calculations on card B was stored in the vram of card A it has to be transported to card B (either by the bridge (don't know if that's possible) or through the pci-e slot, chipset etc), which introduces an unworkable amount of latency

 

[mame]

It's all used, it's just that the ram on each card has to mirror the other one

 

[Judgeneo]

I think both AMD and Nvidia use alternate frame tech for their multiple card setups.

So card one does one frame, then card two does the other. This means the data is more or less mirrored between the two cards.


The reason why they don't combine into one 6Gb setup is most likely bandwidth restrictions.

 

[PorkPieCat]

I think both AMD and Nvidia use alternate frame tech for their multiple card setups.

So card one does one frame, then card two does the other. This means the data is more or less mirrored between the two cards.


The reason why they don't combine into one 6Gb setup is most likely bandwidth restrictions.

this. one card does one frame, other card does the next frame. this also happens to introduce a fair bit of frame latency, but that's unavoidable. anyway the memory buffer for each frame therefore still remains the same

 

[Rroff]

What OptimaLnrg said, the latency of shuffling data over that distance would cause too much of a performance hit - theres a reason the VRAM is usually sitting around the core at equal distance - rather than parked in a nice neat block somewhere further back.

 

[Jarrod]

Wait so, if one card processes one frame, then the other card process the next, rinse and repeat. Does this mean that, the textures for frame one are stored on card one, and textures for frame two are stored on card two, meaning effectively 6GB?

 

[Final8y]

What OptimaLnrg said, the latency of shuffling data over that distance would cause too much of a performance hit - theres a reason the VRAM is usually sitting around the core at equal distance - rather than parked in a nice neat block somewhere further back.

Indeed and you can clearly see that even the length of the individual traces are taken into account.

 

[Addicted]

Wait so, if one card processes one frame, then the other card process the next, rinse and repeat. Does this mean that, the textures for frame one are stored on card one, and textures for frame two are stored on card two, meaning effectively 6GB?

No. All textures are stored on both cards.

Think of it as mirrored RAID, both data on both cards.

Hence why only 3gb is available...

 

[Jarrod]

Ah ok Always wondered and now I have my answer

Follow up question, what exactly is the Crossfire bridge used for? Or SLI bridge for that matter

 

[andybird123]

actually, it's game dependent, some methods of SLI/crossfire involve one card drawing the top half of the frame and the other card drawing the bottom half

in any case, as mentioned above, both cards need to store all of the textures being used because those textures can be needed by either card at any time - cards will even deliberately not clear out used textures (if they have the space) in case they are needed again a few seconds later

a bridge allows the 2 cards to communicate with out needing to rely entirely on the PCIe bandwidth, which is part of the reason why pcie speed and type (e.g. x16 vs. x8 and v2 and v3) doesn't make a big difference until you get to 3 or even 4 card setups

 

[PorkPieCat]

actually, it's game dependent, some methods of SLI/crossfire involve one card drawing the top half of the frame and the other card drawing the bottom half

in any case, as mentioned above, both cards need to store all of the textures being used because those textures can be needed by either card at any time - cards will even deliberately not clear out used textures (if they have the space) in case they are needed again a few seconds later

a bridge allows the 2 cards to communicate with out needing to rely entirely on the PCIe bandwidth, which is part of the reason why pcie speed and type (e.g. x16 vs. x8 and v2 and v3) doesn't make a big difference until you get to 3 or even 4 card setups

thanks, that was informative.

 

[andybird123]

you're welcome