Printable View
Well I was just looking at my PC and thought... SLi works by combining two graphics cards to essentially give you added graphics horsepower!
But it doesn't mean 2GB of graphics on one card + 2GB that's just still 2GB and doesn't mean 4GB altogether... what if the memory was raided? Like storage RAID?
Wouldn't that double the rate of transfer plus also give you more graphics memory? Just a thought... might be possible? Has this been implemented anywhere?
Posted via CB10
they generally arent the same speed at all so you would actually be slowing stuff down
How? Can you elaborate? If there's two mediums and two separate pipes going to each it would be faster? Imagine RAID 0 but with graphics cards, we really don't need one card to handle half the frame while the other does the other half... Just give the whole frame to one card and give the next to the other with a raid controller switching back and forth making it sync...
GPU architecture is already massively parallel, but only very good at performing certain tasks. You can't really think of it as storage RAID arrays (especially RAID 0), as access to within-process data in GPUs is very limited by design. The slow down comes when the much slower CPU or controller needs to retrieve the output. So if you have multiple cards to form "a raid array of cards", you would be limited by the controller/cpu controller,.. so you might as well instead just make a GPU with more parallel pipelines.
Posted via CB10
But as they are a type of CPU they do fetch information from the graphics buffer do they not? It completes tasks in realtime and shoves it out of the HDMI or etc. Another separate memory controller could manage the would be 'RAID' on the mediums because SLi already does alternate frame rendering but by adding something in-between the graphics card and the actual CPU it would be possible in theory... It would also be backwards compatible with older cards because it would be inbetween I guess? Just need another bit of hardware to actually manage this but this is getting a bit too far ahead of myself! I am sure its possible unless someone can actually give me a rock solid explanation on that it cannot be done.
Basically what you are thinking of is how a GPU works. It's already a bunch of very specialized processors (i.e. Limited to a few operations when compared to a CPU) that break up the data, and work on each little bit in parallel to perform very menial (shading, interpolation, etc) tasks very quickly. They don't really share within-process data (except for a few exceptions that have been added to make GPUs more usable for performing non-graphic-related applications (simulations, medical imaging processing, etc.))
Posted via CB10
Yes I understand that, but like how a CPU fetches data and process it and outputs it, it does the same! It doesn't need to manage the RAID, I don't even think the GPU will even notice that the data is being broken up between two GPU's because of its lack of it actually being very intelligent. They don't actually need to share data... It could be managed by another chip. Like how older graphics cards used to work, used to have a separate PCI just for the output. So you think it cannot work theoretically?