Ideal CUDA workstation I'm building a CUDA workstation

I’m working on developing hydrologic models using CUDA and don’t want to have to travel to my campus office (I’m a graduate student) everyday. So, I’m building a desktop workstation (on the model of FASTRA, but not exactly the same) to double as my primary computer in my home office.

I have the following stuff already:

Antec 900 Case
EVGA 780i 3-way SLI Motherboard
Intel Xeon X3220 Quad-Core CPU
8GB Geil Esoteria DDR2 4-4-4-12
2x750GB Seagate SATA HDD
1x1000GB Hitachi SATA HDD

What now? I need a PSU and some GPUs to make this think useful. My problem arises from not knowing which GPU would be best for a CUDA machine. I don’t have Tesla money and want to make this work with regular gaming hardware. The debate is between 3x9800GX2 and 3xGTX280 cards. Also, what PSU to power them?

I know the GTX280 has more stream processors per GPU, but the 9800GX2 has two GPUs per card. So, with 6 GPUs the 9800GX2s add up to 768 stream processors; the GTX280s add up to 720 stream processors. But, if CUDA code can’t be run on a GPU attached to a display for more than 5 seconds, does that mean that only two GTX280 GPUs would be full-time CUDA engines while 5 9800GX2 GPUs?

I’m confused!

Forgot to mention that I’ll be running both Linux x64 and Vista Ultimate x64.

You may ask for some volunteers to run your current code on a both GPU types to help decide which gives better speed for your money, if that’s all that matters.
But in general, you want to run for a G200 board… even if it is less bang per buck than the GX2… [which may or may not be true, it’s very close.]

G200 is much superior for CUDA in general. Double the RAM (1GB per GPU not 512MB) may be a huge factor alone. Double the registers, double floating point support, shared atomics, 64 bit ints, more threads in flight… lots of new hardware features.
the G200 is just super-refined for CUDA. Even if you think you don’t need some feature, at least it will be waiting for you if you do need it later.

Unfortunately, I cannot share anything before the final results are published. But, you make a good argument for the GTX280. I also believe the power requirements for 3xGTX280 is less than 3x9800GX2.

Also, I forgot to ask how double precision floating point works with the GTX280. I know it’s supported, but is it very slow?

I would definitely go with the GTX cards, but you’ll either have to do some adapter trickery or use two PSUs in order to get the number of 8-pin and 6-pin PCIe power connectors you need for three cards. I would consider having a standard display card with no power requirement and just having two GPUs instead.

There is one hardware double precision unit per multiprocessor, so you basically get 1/8 the speed of single precision.

I see that PC Power and Cooling makes a supply with enough 8-pin PCIE connectors. It’s big and heavy, though.

Make sure to get a motherboard that has verified support for PCI Express 2.0 transfers - you get MUCH faster data copies to/from the cards. I have a Gigabyte motherboard that should be PCI-E 2.0 AFAIK, and it only does 1.0 speed transfers. A Tyan server board that we have does the full speed transferes, and there’s a whole lot of difference going on wrt/ speed.

Take the GTX280. You can have 301024 = 30720 threads running at a time versus 16768*2=24576 for a 9800GX2.

I placed an order for two EVGA GTX280 SSC 1GB cards last night. They shipped next day air today, so should have them tomorrow. I’m looking foward to getting the new box setup and running. I just might have to throw a game on this thing too just to check out these cards ( and since the system is hooked upto a 30" LCD)…

Thanks for the help.

Yes, I believe that would be a prudent course of action, sir.

Good luck, have fun with the box.