Originally Posted by Brutuz
The thing is though, thanks to the bridge chip, 8 GPU's looks like 4 GPUs to the OS, the bridge chip gets Two GPU's and works the CF through the Hardware for them, then to get QuadFire with two HD4870x2's, it will look like two GPU's to the OS, BIOS, etc, even though its four.
Although this does sound nice, I'm not convinced that this is really an ideal configuration at this time. The way CrossFire currently works (as well as SLI) is using Alternate Frame Rendering. AFR is a very driver dependant way of implementing multi-GPU rendering. What happens is that the graphics engine of a game will send Frame A to the driver to be rendered. Knowing it has 2 GPUs available, it dispatches Frame A to be rendered on GPU0, but meanwhile tells the game that the frame has completed rendering so that it sends Frame B, which is then redirected to GPU1. Essentially you have the next two frames being rendered independently at the same time. This is the reason that some people experience "microstuttering" because Frames A&B finish rendering very close together, but Frames B&C are farther apart. So you would have something like this:
Frame A <10MS> Frame B <30MS> Frame C <10MS> Frame D <30MS> Frame E, and so on.
I don't think there would be an easy way to implement this in hardware, nor do I think there's a practical reason for doing so. The reason I say this is because the only advantage a hardware AFR implementation would have is transparency to the driver, so it would allow more GPUs. However when you add greater numbers of GPUs to AFR mode, frames syncing with actual in game events begins to become an issue since it is essentially "lying" to the graphics engine about when a frame has been rendered. The result is an improportionate difference between the framerate increase and the actual perceived performance benefit.
Now if they really did want to do a multiGPU setup entirely in hardware, the quick and easy way would be to split the frame, have one GPU render one half and the other render the other half, put it back together, and send it along the pipeline. This is also not without its issues, however. In situations where one half the screen is significantly more complex than another (like perhaps if the top half of a scene was mostly sky and the bottom was populated with trees and characters), the performance avantage is going to be very minimal compared to a single card because you're still left with one GPU doing the complex stuff. If ATI really wants to focus on scaling (and I think they do), they're going to have to start working on alternatives to AFR, and this is likely going to mean developing a way to load-balance split-frame rendering so that each GPU finishes its half of the frame at the same time. I'd like to think that R700/X2 is delayed due to drivers
because they're working on a more balanced SFR mode, but the information they've provided so far doesn't say anything one way or the other. However, even if they are working on it now in software, I think seeing a hardware implementation is still far enough away that it wouldn't be in range of the next seies. That being said, nobody outside of ATI really knows how long they could have been working on this so there's really no point in guessing at how far away we are from seeing this.Edited by darkcloud89 - 6/21/08 at 11:09am