Bitcoin Forum

I find various sources on this forum claiming that the AMD linux drivers only support up to 8 GPUs (i.e. max 4 dual GPUs).
I however find a lot of conflicting information on this forum when it comes to the number of GPUs that can be driven on windows (6, 8, 256, ...).

It's physically possible to build a rig with 7x6990, on a standard motherboard with 7 PCI-e slots (of any kind). But will the hardware (address bits?) and the windows drivers be able to handle this? Or will there be other software problems? I couldn't find an example here of a rig running >8 GPUs on the same motherboard, nor could I find any explaining why or why not it would be (im)possible.

Apart from the aforementioned software limitations , you'll run to the problem of having not enough power. How much of a psu and motherboard do you expect to e able to run 8 full fledged gpus?

The newer drivers (11.5+ ?) support 8 GPU in Windows.  Older drivers only supported 4 GPUs under Windows.

32bit OS sometimes have problems (even Linux) with the full amount of GPUs.  Others have reported problems getting full number of GPU under Windows XP.  Using Windows Vista/7 x64 you can get 8 GPU (4x6990) working.

Nobody has gotten 9+ AMD GPUs working on any OS under any conditions and since AMD shows no interest in becoming a real GPGPU competitor likely never will.

When not using a case, it's not so difficult to use multiple PSUs, and with PCI-e extender cables with molex connector, this seems fairly doable, since the power won't go through the motherboard. Or am I wrong in this?

That is a non-issue.  Dual PSUs and powered extenders solve it easily.  The driver, OS, and BIOS limitations are essentially unsolvable though.

Would it be a possible solution to mix vendors (say 4x dual AMD, 3x dual nVidia)? Or is it out of the question to ever achieve this in any reasonable way (and should I better forget about running dual cards in this way)?

It seems stupid that there are many boards with >4 GPU slots if >4 dual GPUs is impossible. :/

Mixing AMD & NVidia is a recipe for disaster.  You "could" (with custom BIOS, and Linux kernel) use 9+ NVidia GPUs on a single board but they are worthless for mining.

Honestly there isn't that much economic value to going >8.

A motherboard, USB drive, 2GB of RAM, and Sempron are dirt cheap.  Maybe $200.

So going 8 GPU per board you are paying $200 overhead per board or $12.50 per GPU.  Say you figured out a way to get 16 GPU on a board.   Your overhead would be $6.25 per GPU.  You "saved" a whole $6.25 per GPU.

Now lets say you built a 100 GPU farm.  Your savings would be $600 (minus the cost of more expensive 8 slot boards, and any work requires to get 9+ GPU working).

When you consider the amount of cost, planning, wiring, cooling, maintenance, and troubleshooting goes into a 100 GPU farm that savings is a drop in the bucket.

It definitely is a viable solution, except implementing it will be far from trivial.
Turn off any unused device in the BIOS for should you run out of resources at the BIOS level, there is nothing the OS can do - some GPUs will remain uninitialized / unmapped into memory.
Far easier to use two systems with 4 cards each.

Would you agree there is a distinct possibility that boards with multiple GPU slots may have been designed with maximum flexibility in mind?

Oops. Missed that completely  ;D

currently using semprom 1 gig ram 2gb usb 2 ati gpu combo so far it is the most price/mh efficient solution , a bit heavy on the power though