Multi GPU shared memory

[EliteRetard]

They make and use fake PCI-E chips or whatever to connect two GPUs on one PCB...why don't they make a chip that also relays memory commands and such so they can share the installed memory? You could use less chips which would take less space, use less power, and lower heat, potentially making it lower cost as well. Possibly improving performance, since sharing resources is faster than each side working on a separate piece of the same thing. If the cards can share what they're doing, and then send the completed work to a shared buffer to be released would that eliminate micro-stutter?

Instead of the current standard example 2x 2GB cards (4GB total) but only 2GB effective, it'd be 2 cards sharing 2GB for the same effective 2GB. Or using the 2x 2GB and sharing that for 4GB effective (ideal solution?).

Don't know any of the technical stuff, I'm just throwing this out there...like one of those "hey why didn't I think of that" kind of things. Has anyone tried this before, or are working on it currently? Is there some reason this wouldn't work?

 

[digitaldurandal]

Latency

 

[EliteRetard]

At 5+ GHz and 384bit busses, how much would latency have an effect? 1%? Or more significant like 10%?

 

[EliteRetard]

Ran into my old thread, I still wonder about this. Does anybody have additional input this time?

 

[Grooveriding]

Each card takes a turn rendering a frame afaik. So each card is rendering a unique frame and using its memory to do so. I'm not really clear on the specifics of how it works, but I would guess sharing the memory wouldn't be as effective as each card using its own onboard memory.

The latency suggestion also makes a lot of sense as the cards would be passing information across the bus & bridge to access the other card's memory and that information would also have to come back.

 

[BrightCandle]

In Alternate frame rendering each of the two GPUs is rendering a different frame. Thus they have quite a lot of differences in the actual view port and state of the game. So while the textures are the same much of the rest is different. Sharing wouldn't save very much.

Worse than that is you would need an interconnect that was ram speeds, along with some mechanism to resolve concurrent access. And the chips now need to worry about their cache being out of date. Even Intel on the same silicon can't get its CPUs to speak quickly enough for cache coherence, and here we are talking about cache coherence across different chips. The memory at 400GB/s is going to be a challenging interconnect to begin with.

For all intense purposes its not going to net much gain, be technically very difficult to do and only help on dual card solutions to save a few tens of dollars in RAM.