Has anybody experince with installing four 8800GTX cards on a single mainboard like the P6N Diamond from MSI ? Can you handle each device independently by cuda ?
I’ve got 3 GeForce 8800GTX cards in an Asus motherboard on a Linux machine and they all run CUDA threads concurrently… I assume it’s doable on Windows as well.
If you are using the air cooled 8800 boards, you will not be able to fit 4 of them into that motherboard. The last slot does not have enough space next to it for the heatsink and vent. If you get the water cooled version of the 8800, then it might work since those are single slot boards, but it will cost you an extra $300 each. You can also wait for future single slot 8xxx cards which won’t require water cooling. MSI showed a few at CeBIT, though they are likely to be lower performance than the 8800.
(As a general note, you are probably not going to find any ATX motherboard that can take 4 dual-slot cards. The ATX standard only defines up to 7 card slots.)
Please, could you say what exactly the Asus motherboard you have?
We are thinking install 2 cards
The motherboard John is talking about is an Asus P5N32-E SLI.
Yes, that is right. We would also like to hear what other motherboards people test out, particularly boards with multiple x16 slots like the Asus boards we’re using currently.
Thanks mfatica and John
John, which power supply are you using on your 3-card setup?
We have a couple different ones in use currently…
It’s quite easy to setup a machine with 2 GPUs, doing 3 is an exercise in balancing the power load and coming up with 6 PCIe power connectors.
A very good choice is the Thermaltake 1200 watt PSU, which I recently put into one of the test machines. The other one has an 850 watt PSU. The 1200 watt PSU is certainly the easier of the two configurations to make workable and safe for sustained use. In practice our two 3-GPU test machines are only drawing around 700 watts with all 4 CPU cores and all three GPUs running flat out. The main issue is having a PSU that provides enough amps and has enough connectors for all three GPUs. With the 1200 watt Thermaltake it was quite easy to setup. The first 3-GPU machine using the 850 watt PSU was a bit more of a hack job and I wouldn’t bother going that route if you can just get a 1200 watt PSU to begin with. The thermaltake 1200 watt PSU also has the new 8 conductor PCIe power connectors, which are apparently going to be used for some next generation GPUs, but it also comes with converters to use them with existing cards that use the common 6 conductor PCIe power connectors.
Anyway, the main issue is having a PSU with enough amps and enough PCIe connectors, so long as you manage those to issues it’s not too hard.
I’ve got a machine running windows setup with two cards, and yeah power supply and such is easy, but I’m having trouble with windows reading the secondary card not being used for display. Any idea how to skirt that issue? I’m running an 8800 GTX and an 8600GT (display) on an XPS.
Regarding of setting two different cards, perhaps you can check this thread: http://forums.nvidia.com/index.php?showtopic=29567
Hope it helps a bit.
Are there motherboards that fit 4 NVIDIA 8800 GTX, that can be used for computing together with cuda? As I see the Asus P5N32-E SLI fits 3 of these cards.
I am new to this arena and I am currently setting up a new PC to do financial simulations and pricing. I bought a X38-based machine (which does not do SLI) and I want to install two GeForce 8800GTX video cards in them. Can I do this eventhough SLI is not supported and still use it to play games (without SLI of course) and use CUDA on both cards?
CUDA will not benefit from SLIing cards together. So if your primary goal is CUDA then just having two cards without SLI is okay. You will be able to use one the cards to play games =)
The X38 does not support SLI and is a ATI crossfire only board. I am planning on installing two 8800GTX on these boards without the SLI connector and specifically turn off SLI on the driver. That should be OK right?
For everybody how wants to use more than 2 cards on a single board:
It is important for scientific computing applications using domain decomposition and communicating over boarders that for every GPU there is at least one CPU- core.
So you need a quad core if you are using 3 or 4 GPU cards. The reaseon (tachyon_john told me) that the thread synchronization is cpu intensive…