Why has stacking vRAM been so difficult?

[TechGod]

As the title says. 

[79wjd]

Because the way SLI/XFIRE work, both GPUs require the same information in memory as they both work together rendering the same scene (and there would be far too much latency for cross card memory access to occur).

 

/VRAM stacking is only an issue with SLI/XFIRE, not multi gpu configs in general. 

[TechGod]

Because the way SLI/XFIRE work, both GPUs require the same information in memory as they both work together rendering the same scene (and there would be far too much latency for cross card memory access to occur).

 

/VRAM stacking is only an issue with SLI/XFIRE, not multi gpu configs in general. 

Would some sort of proprietary connector that is faster than PCIE 3.0 be able to be used to connect the GPUs together to reduce latency? 

[QueenDemetria]

As the title says. 

Because after the 5th chip they topple over.

 

jk, but CFX/SLI function in the sense that the cards work together, but they don't share resources. A gpu will render a frame, then the other GPU renders the next frame, the this cycle repeats. This is called "Alternate Frame Rendering"(AFR).

[TechGod]

Because after the 5th chip they topple over.

 

jk, but CFX/SLI function in the sense that the cards work together, but they don't share resources. A gpu will render a frame, then the other GPU renders the next frame, the this cycle repeats. This is called "Alternate Frame Rendering"(AFR).

Right, but as I posted above, can't you use a faster connection so the two GPUs can work on two different things. I mean right now, it sounds like a total waste to have multiple cards in a gaming rig. 

[mikat]

Would some sort of proprietary connector that is faster than PCIE 3.0 be able to be used to connect the GPUs together to reduce latency? 

not likely, the amount of bandwidth from the gpu to the vram is so much higher than any connector we have right now, that we will have to rely on DX12 to solve our problems

[TechGod]

not likely, the amount of bandwidth from the gpu to the vram is so much higher than any connector we have right now, that we will have to rely on DX12 to solve our problems

even faster than the proposed PCIE 4.0?

[79wjd]

even faster than the proposed PCIE 4.0?

yes. 

 

Think about the reasons behind HBM -- i.e. getting the memory as close to the core has possible to improve performance. There is a physical limitation on connection speeds, and distance actually plays a crucial factor.

[Dabombinable]

not likely, the amount of bandwidth from the gpu to the vram is so much higher than any connector we have right now, that we will have to rely on DX12 to solve our problems

TBH, Nvidia needs to start using a new connector since their SLI bridge is 9Gbps (that's if I'm remembering it correctly-either way its slow).

 

yes. 

 

Think about the reasons behind HBM -- i.e. getting the memory as close to the core has possible to improve performance. There is a physical limitation on connection speeds, and distance actually plays a crucial factor.

Though the shorter distance is mainly to help reduce latency, with the connection material, resistance and width (as in width of the track) affecting speed.

[BrightCandle]

The speed of electricity is an issue at these sorts of speeds. Even a very wide connector would be 100,000x slower in latency than going to vram on device and transferring in the order of 500GB/s would be extremely difficult.

[TechGod]

That's insane data that needs to be transferred. Thanks for all the replies