Crossover/SLi Memory

[Dezz]

Hey,

 

can anyone explain me how exactly does the memory of cross/sli GPUs works? The little i know is that if the cards are 4gb each, cross/sli is still 4gb not 8gb. What happens with the rest of the memory, memory of which card is being used (mmeory of one card, or is uses small portions of memory from each card?)

 

Also the dual GPUs like Titan Z, 295x2, Ares III, why do they have for example 8GB (4GB for each chip) when they can actually use only that 4GB? Why don't they just simply remove the memory that can not be used? Is it because of marketing or were they planning to somehow merge the memory so it can be used togather as 8gb?

 

Please dedicate me into this

[Archangel1994]

Because they use alternate frame rendering. First a frame is done by GPU1, then the next is by GPU2, then by GPU1 again and so son. The GPU's cannot access the other GPU's memory

[Fate]

I believe the memory usage is mirrored between the GPU's.

In the future Mantle/DX12 may be able to use all of the memory in crossfire/SLI, but we'll have to wait and see.

[zappian]

The memory is mirrored .

 

If you read a post here on LTT you will se that mantle and DX 12 will use the fullmemory buffer on cfx probably.

 

EDIT: I swear i didn't read your comment fate but we said the same thing

Great minds think alike

[SkilledRebuilds]

Because they use alternate frame rendering. First a frame is done by GPU1, then the next is by GPU2, then by GPU1 again and so son. The GPU's cannot access the other GPU's memory

@Dezz - ^ One of the easiest to understand explanations.

[Enderman]

When using alternate frame rendering, each GPU needs to render every second frame. The PCIe lane bandwidth is only about 30GBps while the vram on a GPU is hundreds of GBps.

This is why the frames need to be copied on each GPU, and each one takes every second frame.

 

One solution is to remove every second non-used frame from each GPU, so that the GPUs have alternate frames in their vram buffer, giving a total of 4+4GB instead of just 4The problem with this is that the GPUs dont sync up correctly to give frames at the right times, but maybe AMD is working to fix this and make 4+4GB CFX possible.

[Dezz]

Because they use alternate frame rendering. First a frame is done by GPU1, then the next is by GPU2, then by GPU1 again and so son. The GPU's cannot access the other GPU's memory

 

Thank you, so basically they can use all of their memory, 8gb in this case.

 

Would it be possible that both of the cards could access "one" 8gb memeory? (there would be one 8gb memory for both, not 4gb for each). Therefore if game wouldn't support 2GPUs, that one GPU could use fully 8gb.

 

 

EDIT: Sorry for stupid questions, I'm curious

[Archangel1994]

Thank you, so basically they can use all of their memory, 8gb in this case.

Yes they can, but each GPU can only access 4GB. The other GPU accesses a different 4GB

 

 

Would it be possible that both of the cards could access "one" 8gb memeory? (there would be one 8gb memory for both, not 4gb for each). Therefore if game wouldn't support 2GPUs, that one GPU could use fully 8gb.

That will happen with Mantle/DX12/Other low-level API's. They will use split-frame rendering, both GPU's working together on the same frame instead of alternating who renders the frame. This effectively means a full 8GB

[Dezz]

Yes they can, but each GPU can only access 4GB. The other GPU accesses a different 4GB

 

 

That will happen with Mantle/DX12/Other low-level API's. They will use split-frame rendering, both GPU's working together on the same frame instead of alternating who renders the frame. This effectively means a full 8GB

 

Thank you, smarter every day  B)