GTX285 vs TeslaC1060 vs FX5800 hardware question

Accelerated Computing CUDA CUDA Programming and Performance
1

I was sure that there were multiple threads for this already, but for some reason I’m not finding in the archives. Apologies for duplication:

I’m parting out a new workstation for CUDA use, and I just wanted to clarify a couple things. First off, the GTX285, TeslaC1060, and FX5800 all use the same GPU, correct? And the primary end difference being 1-2GB ram on the 285, and 4GB on the Tesla and FX cards? Memory aside, are there any real-world performance comparisons between the cards?

As it stands now, I can’t see any reason for spending ~$3000 for an FX5800 or ~$1500 for a C1060, when a $400 GTX285 seems like it’ll give me the same performance (provided of course I’m fitting everything into the ram.) Could someone point out what I’m overlooking?

Thanks!

2

The burn-in tests on Tesla and Quadro. Other than that, GTX285 will have the best performance if everything fits into RAM.

3

Thank you, this was my understanding as well. So keep an eye on temps, and get a case with good airflow, and no worries?

4

Yeah, that should be fine. Just make sure you get a fairly large power supply as well, since multiple high-end GPU’s will draw some serious power.

As for the burn-in tests, I think E.D. meant that they actually run the GPU’s for those cards at full speed/power for 24 hours (or some other length of time) to make sure there are no defects in the chip. This is probably an expensive process and unnecessary for normal users, so they only do it for the ‘pro’ cards.

EDIT: I believe the Tesla 50% sale (for developers) has been extended until the end of this month…so that makes a big difference in price if you think you’ll use the memory eventually.

5

Right, and stay away from overclocked cards. It’s also worth running memtestG80 on the card when you get it:

http://folding.stanford.edu/English/DownloadUtils

as your own “memory burn-in” test.

6

Yeah, but I’m sure you’re aware the wheels of academia can turn slowly. Highly unlikely we’d have everything approved by the end of the month :)

My current thought for a build consists of:

ASUS P6T LGA 1366 Intel X58 ATX Intel Motherboard (has 2 PCIe-x16 2.0 slots)

EVGA 02G-P3-1185-AR GeForce GTX 285 2GB (for computing)

XFX PVT94GZAH2 GeForce 9400 GT 1GB (for display)

Intel Core i7 920 Nehalem 2.66GHz

along with a 1000W PSU and 6gb ram, dual booting linux64 and vista64. Just putting that out there in case there’s a glaring incompatibility.

7

Unless you’re on a really tight budget, why not go for 12GB of RAM? It’s not that much more for another 3x2gb kit ($100-150 probably) and it may come in handy if you’re working with large datasets (if you have to swap several datasets back and forth to the card, it will eat up your 6gb in a hurry since some will already be used by the OS, etc.)

8

Quadro series cards are optimized for professional applications (like CAD or 3D) offering HW implementation of features used in these applications (e.g. line antialiasing). Therefore, for CUDA apps I would choose either GeForce or Tesla card. As guys mentioned, GeForce’s weakness is power consumption and, of course, high temperature. Anyway, we have bought GTX285 for CUDA purposes and it seems to work fine. :)