I remember there being a limited number of frame buffers available in DX9, which caused multiple-GPU solutions like 7950GX2 in SLI to not be fully functional. (one GPU would simply remain idle as there was no fourth frame buffer to utilize it) With the impending release of the 9800GX2, "quad" SLI is again an option. Could the DX10 architecture provide utilization of all four GPUs in a system with two 9800GX2s?

Update bump:

I called Microsoft and to my dismay, I was shunted between a few departments until I got to one person who told me that I would need to contact the manufacturer of the card in question to find out my answer. When I explained that I knew everything I wanted to know about the card and that I had a question about a Microsoft product and not about an nVidia product, she repeated herself. When I asked why she thought that nVidia would know more about DX10 than Microsoft, she again repeated her tired "you're going to need to contact another company about one of our products" refrain. When I asked her if I was being given the brush-off because Microsoft was trying to hide something about DX10, she hung up on me.

*clap clap* Excellent customer support, guys. Really good job of showcasing the product.

Wow that really is quite bad.... and now I am going to 'shunt' you off elsewhere too ;-)

Check out this FAQ for some better places to get native DX support - sadly these forums don't get the traffic, or the MS employees to answer complex questiosn like yours. I would suggest the directx-l mailing list in this case.

 

Part of the problem here is that the fact that you're running on multiple GPUs is not visible at the API level. Currently I believe you have to use some hardware specific APIs to find out how many cards you're actually running on, and even then the API won't do anything with that information.

Most multi-GPU systems just interleave frames on different cards, so from an API standpoint, the only thing that matters is that you have enough frames queued up to use all of your GPUs (for obvious reasons, you'll want to avoid any cross-frame dependencies here). There is no explicit scheduling of jobs across the different GPUs, that is all left up to the driver.

So to answer your question about frame buffers, I honestly couldn't tell you 100%, since I don't work on the HW drivers, but I would presume that it isn't a problem, since the card that is doing the work will always have access to its own local memory.

John, the problem with SLI is the pre render limit.

If you want to feed 4 GPUs you need a pre render limit of 4. But the default is only 3. Therefore the runtime will queue 3 frames and 3 GPUs can start working. But there will be no data for the 4^th GPU in the queue.

I know that Direct3D9 Ex has a method that looks like that it can change the pre render limit to up to 20. But I never tested it. On the other hand I wasn’t able to find any information about the pre render behavior of Direct3D 10. The only think that seems to be sure is that there isn’t any API to change the pre render limit.

Thanks for the help, all. I was real surprised to find three people hit the thread so quickly with helpful replies.

I think demirug has hit the nail on the head here... I wonder if people with 3870X2s in Crossfire are noticing this yet? Hopefully someone will work out a way of increasing that render limit for users of multiple-GPU cards. (not a big market, I know, but one that's full of tweakers, so there's some hope)

I'm getting outside of my area of expertise here, but can't you just create a swap chain with 4 buffers?

MrMonroe,  see this GDC 2008 presentation.  I havent read it, but its talking about using multiple GPU's.  (Thanks Zman for pointing these out).

http://developer.download.nvidia.com/presentations/2008/GDC/GDC08_AD3D_MGPU.pdf