Motherboard for GPGPU

[ismailfaruqi]

Hello, currently I'm building a workstation for GPGPU research. To maximize the number of computing units, I'm looking for a motherboard that has four evenly spaced x16 slots, and preferably has also built-in video chip, as we try to free compute unit memory from being allocated as frame buffer.

It can be Intel or AMD board.

Till this moment, the only motherboard that qualifies my requirement is MSI K9A2 (and its V2 version), but unfortunately it doesnt have built-in video chip.

Anyone knows other alternatives?

Thank you very much.

 

[MegaWorks]

I bought the ASUS M3A32-MVP Deluxe, using 3 9800GT but the fourth slot won't take my last 9800GT.

MSI K9A2 Platinum That's a very good price for a high end board.

ASUS M3A79-T Deluxe the new replacement with a better south bridge chipset.

I would take the MSI and pair it with the new Phenom II processors.

 

[MegaWorks]

Oppss I didn't catch your last sentience. You don't need built-in video chip when you're GPGPU, you can use any normal card to view the monitor.

 

[VirtualLarry]

Yeah, I built a Folding@Home box with four 9600GSOs, on a MSI K9A2 Platinum, probably a v1.0. Has four PCI-E slots, fairly decent price.

Check out link

There are only a few cases on the market that support the additional slots in the back, in order to use four double-width graphics cared. Some Lian-Li, and the ThermalTake Armor.

You might want to consider that too, if you plan to use double-wide graphics cards.

Mine are a mix of double- and single-wide graphics cards, so I put mine in an Antec 300. Video card temps are still through the roof, even with five fans on high. Top card was running at 96C, which is still lower than the 100C that Fastra runs at, but my card died after two weeks of 24/7 folding at that temp.

Edit: If you run Folding, then use the 180.xx Nvidia drivers. They reduced the CPU requirements under Windows XP to virtually zero.

 

[aka1nas]

Originally posted by: MegaWorks
Oppss I didn't catch your last sentience. You don't need built-in video chip when you're GPGPU, you can use any normal card to view the monitor.

You do for CUDA, otherwise there is no guarantee that a resolution switch will not overwrite the CUDA kernel if it needs to allocate more framebuffer memory (at least as of a few months ago). It's probably an issue if he is running parallel algorithms and needs to manage memory across all 4 GPUs.