VRAM bandwidth in Crossfire/SLI?

I'm curious about something here. When you run graphics cards in Crossfire or SLI, they obviously mirror their memory, so two cards with 2 GB still end up with 2 GB of usable memory.

But how does it work regarding bandwidth? If each card has, say, 192 GB/s memory bandwidth, can they still independently access their memory and thus yield an aggregate bandwidth of 384 GB/s, or does it again get limited to what a single card would have?
 

stotty5

Distinguished
Jan 20, 2012
4
1
18,510
When you are running SLI/Crossfire, you are sharing 2 resources on the system. Each card should handle half of the work (how depends on the rendering method) with their respective processor and share the same 2GB of memory maximum, but the advantage presented here is that the memory bandwidth is shared between the cards allowing double the available bandwidth when reading memory of a single card. Since the 2GB is shared by the two cards, the writes are still limited to a single card bandwidth, but this should be limited in gaming scenarios.
 

Right, of course the writes would have to be limited to half bandwidth because to keep the same data in both sets of memory you obviously have to duplicate all writes.
But reads are indeed what it comes down to. Unfortunate that we seem to have a split decision so far. :(