I am currently looking for nvidia graphiccards to plug them into machines that students can use here at heilbronn university. Due to the fact we want to provide “a first contact” with Cuda, I think we don’t need the most powerful cards out of the G200 series. I thought a GeForce 8800 GT(s) would be enough to get in touch with Cuda. These cards can be bought cheap nowadays.
What do you think of it, or do you think it’s better to buy the 9000 series?
Thanks in advance,
9800GX2 provides best price/performance ration, I think.
G200 might be better for education because it supports double precision and makes memory coolescing easier. If you don’t need this, 8800GT should be fine – I’m using 8800GT+8600GTS in my development machine and haven’t seen much problem.
I bought an 8800 GT with 512MB memory a few months back…only cost something like $130 from Newegg. However, I imagine you will want these cards to be a long-term investment, and so I will agree with Andrei and say that you should go with one of the newer double-precision cards. It won’t cost too much more up front, and since you’ll have the cards for a while (at least a few years, I imagine) it would make better sense to get cards that your students will probably be using “out in the real world”.
I would advice you to get 2 Tesla 10 series PCI-E cards. It will remain an asset to the university! + 1 cheap NVIDIA graphics card so that main display remains on the cheap card while the TESLA is dedicated for computation. This way, you will avoid the watchdog timer issue as well (with the latest drivers).
Thank you all very much for your hints and thoughts!
I have to buy these cards for 5 machines, so the Tesla solution is a bit too expensive. I checked the prices for the GTX260 and I think I’ll buy these cards for my machines here.
Think about making one or two machines have dual cards. It’s an extra skill to program CUDA in multi-gpu, and a useful one since the most powerful cards (like 9800 GX2 and upcoming GTX295 X2) are dual internally.