If I buy two cards for CUDA computation two gpu cards

[GoGdizzY]

Is there any restrictions, e.g. to Motherboard? I saw in WIKI:

Nvidia SLI and ATI Crossfire

Nvidia SLI and ATI Crossfire technology allows two or more of the same series graphics cards to be linked together to allow faster graphics-processing capabilities. Almost all medium- to high-end Nvidia cards and most high-end ATI cards support the technology.

They both require compatible motherboards. There is an obvious need for 2x PCI-E 16x slots to allow two cards to be inserted into the computer. The same function can be achieved in 650i motherboards by NVIDIA, with a pair of x8 slots. Originally, tri-Crossfire was achieved at 8x speeds with two 16x slots and one 8x slot; albeit at a slower speed. ATI opened the technology up to Intel in 2006, and all new Intel chipsets now support Crossfire.

SLI is a little more proprietary in its needs. It requires a motherboard with Nvidia’s own NForce chipset series to allow it to run (exception: select Intel X58 chipset based motherboards).

It is important to note that SLI and Crossfire will not usually scale to 2x the performance of a single card when using a dual setup. They also do not double the effective amount of VRAM or memory bandwidth.

I don’t know if this issue was old, and I want to hear about some experience for many-cards COMPUTATION, e.g. it seems impossible to exchange data between the cards?

[seibert]

CUDA does not use SLI in any way, so there is no restriction on the motherboard you install the cards into. (I have two cards installed in a motherboard using AMD’s chipset.) In fact, prior to CUDA 2.3 (which is in private beta now), you have to turn SLI off in order for CUDA to see both cards.

Programming multiple cards with CUDA requires that you run one host thread per card, and partition your workload between them manually. Data cannot be transfered directly between cards, only between a card and the host CPU.

[GoGdizzY]

Thanks！ External Media

CUDA does not use SLI in any way, so there is no restriction on the motherboard you install the cards into. (I have two cards installed in a motherboard using AMD’s chipset.) In fact, prior to CUDA 2.3 (which is in private beta now), you have to turn SLI off in order for CUDA to see both cards.

Programming multiple cards with CUDA requires that you run one host thread per card, and partition your workload between them manually. Data cannot be transfered directly between cards, only between a card and the host CPU.