Good computer for CUDA development How to choose the components?

Hello folks,

My university is going to buy a new computer exclusively for CUDA development. I've recommended to buy a GTX 280 (the Tesla cards a little too expensive, for now  :smile: ).

The problems is, how can we choose the other components of the computer in order to avoid problems? I’ve read in a few topics that certain motherboards can give issues, and the like… there are some important information I should be aware of before proceeding to buy this computer?

Thanks a lot.

Try to buy a computer from a vendor that has a GTX280 in it, then you don’t have to worry. Otherwise :

  • Motherboard that has PCI-Express V2
  • Big power supply that has both 6 & 8 pin power connectors for PCI-E

You already found prices for the new tesla’s?

I wonder if it is possible to power up GTX 280 from the power supply that has only one 6-pin connector for video card.

Power supply is FSP’s OPS600-80GLN, 600 Watts. Are there any adapters from standard molex to 6-8 pin connectors ?

from 6 pin -> 6 & 8 pin? I think that will not work. I believe I read somewhere that if you have 2 6 pin connector, you may use an adapter to make 1 6 -> 1 8, but I believe you can blow out your power supply like that.

No, not from 6pin to 6+8 pin, but from molex to 8 pin.

I have a 600W PSU, but it’s a couple years old so it only has two 6 pin connectors. The eVGA card I got came with both
6+6 -> 8 and molex+molex -> 6
adapters. Unhappily, I crossed my fingers and tried it, and all was well.

On, look at the PICTURES of the cards, which have a shot of the included cables and such, you can see which brands include adapters and which ones. I picked the eVGA mostly because I knew ahead of time I was in a PSU grey zone, my gamble worked.

I’m disappointed in the eVGA card, though, it proudly talks about thier GX 280 color power LED for diagnosing any power problems… the GX 260 does not have it. Except when I got my card… the GX 280 doesn’t have it either. Sigh. I’d love it if board level probes could tell you voltage/amps being delivered by each connector.

Since my board is running CUDA programs for over 24 hours straight, I suspect my power will be fine.

Back to the topic of futureproofing your development platform:

Get a motherboard with TWO x16 PCI-E slots with DOUBLE SPACING. Even though you’ll get just one card now, and only need one slot with double spacing. It will be very tempting later to pop in a second card, and if your motherboard can’t fit it, it will cramp your options considerably.
Look for one with TRIPLE double-spaced slots for even more flexibility.
There is at least one known MB with quad double-space slots, but that may be one-of-a-kind.

And get 4GB of RAM at least, and use a 64 bit OS (Linux preferred, but XP is OK), and a quadcore CPU.

Choose a power supply from NVIDIA’s list of SLI certified components here:

That will ensure that you have the power and connectors needed for your configuration.

Try to match the RAM to the motherboard you choose. Motherboard manufacturers sometimes list on their support websites RAM parts that tested well with each model. User reports in newegg and in the motherboard manufacturers’ “community” forums can also be helpful in that regard.

A larger chassis is easier to build a system in.

Those are some of the things I learned in putting a CUDA box together recently.