Originally Posted by HardBall
Yes, and the total bandwidth is also 2X256bit at 512.
Would it do you any good? Absolutely not, since no single GPU is fast enough to use close to 256 MB of memory for frame buffer, so no two GPUs will be fast enough to use 512 MB. The GPU has to speed up quite a bit to make use of 512 MB of memory effectively. The main benefit of the SLI is to share the load between the two GPUs. Today, video cards generally have too much memory to be economically efficient, with very miniscule improvement from 128 to 256.
Ummm NO! first of all nope sorry it dosnt work that way 2 cards with 256 megs of ram in SLI gives you a max of 256 megs of buffer sorry. Secondly ummm the speed of the GPU and the size of the frame buffer arnt linked the frame buffer mostly stores textures there are no games out there that realy use 256 megs worth of buffer, even the realy intense games are just starting to use more than 64 megs for the most part, which is why under most conditions a 64 meg card will perform just as nicely as a 256 meg card, sticking all that ram onto a card is mostly a gimick to attract people who know very little about computers. I will admit that we are to the point where 128 meg cards start to show a real advantage at high res with AA and AF in newer games, but 256 is still pretty much pointless. Anyways the size of the frame buffer dictates how much texture information can be stored there before the card has to use the AGP or PCIe bus to snag more info at a slower rate, the width of the memory interface and the speed of the ram determines how much faster the card can get needed data from it's frame buffer to it's GPU for processing. Now as far as SLI is concerned both cards dont work "togther" you basicly have 2 seperate 256 meg frame buffers, both working at their normal speeds each feeding data to a seperate GPU, the 2 cards then split up the rendering tasks, there are several easy ways to do this there is Alternate frame rendering, where one card renders frame 1, the other card renders frame 2, then the first card renders frame 3, ect. There is line interleaving where the cards each render an alternate line on the screen each one pixel wide. There are also systems that split the screen up into a grid of some sort, the simplest is top half bottom half, some of the more complicated break the screen up into rows and colums, this is AFAIK the method Quantum used on their high end 8 and 16 way 3dFX based boards, There are some high end ATI FireGL cards out there that work the same way, I think they are 4 way currently. I also belive that there were some attempts to work up an active system where the cards were actualy spliting up non fixed areas based on the polygons that were being rendered I dont know that anything ever realy came from that. SLI is not the best thing since sliced bread, if you dont belive me ask 3DFX about the problems they had with the Voodoo's Or ask ATI about their MAXX line of cards, or Even Volari. Honestly Nvidia's SLI is limited by the 16X PCIe, so basicly you have 2 cards runing in 8X PCIe which is why Gigabyte already has 6600 SLI on one card, and you can not run 2 of those boards together. Geting the cards to "work" together is often a nightmare, it can lead to flickering displays, or bad tearing, or a de interlacing effect, and they have to talk to eachother and comunicate so that each card is geting the right share of the data, and when that dosnt happen they usualy fall back to the master card doing all the work and the slaved card doing nothing which is why a lot of the reviews raise vaild complaints about the fact that in many games there is no performance gain at all. So in summation SLI gives you a max 256 meg frame buffer, that dosnt work any faster than normal, however each of them has less work to do about 1/2 as much in fact, assuming that the cards can communicate properly and spit the work effectively. Some of the ways to render like alternate frame rendering which is BTW the easiest to do, since each card is rendering a full frame you dont get any benifit at all to maximum texture storage. In the more advanced methods of work distribution I suppose there is some benifit because each card is only rendering part of the texture data that is displayed on the screen at one time.