gonna buy a CUDA card tomorrow - Which one?

I got the OK from my advisor to spend about $200 for a card to get started on. I was just gonna get the 8800GTS from newegg or something, but if anyone knows a better card that I should get, I’d love to know. If we get some nice results, we’ll probably spend bigger bucks later, so it will mostly be a learning/experimenting card.


The GTX 260 is probably the best card close to that price range, it will do also double precision.

Determine what version of PCI express is in your computer…if it’s an older computer (<= 2006-ish), it might have PCI Express x16 (not PCI Express 2.0 x16). Mine has the older one, and I just got an 8800GT card that I’m pretty happy with so far (for less than $200).

If you’ve got the newer slot though, get the GTX260 for sure. The double precision will be great to have for your research.

PCIe 2.0 cards are backwards compatible with PCIe 1.0 slots, just as PCIe 1.0 cards work in PCIe 2.0 slots. Buying less than a GT200 at this point is kind of silly, considering how useful the CUDA improvements on that chip are.

(believe me, I’ve done this quite a bit. 8800 GT is PCIe 2.0 anyway :P )

You can use a GTX in even older PCI slots. You won’t get as much bandwidth as a modern slot but for most applications it won’t matter.

I have my 280GTX running in an 8x PCIE 1.0 slot right now.

As for board recommendations, yes, GTX260. The G200 class hardware has huge CUDA improvements over G80/G90! GTX260 is a great sweetspot.

The GTX260 is 1.3 HW, which means for CUDA that it’s much, much easier to exploit memory bandwidth in kernels, as the coalescing hardware is just much, much better.

Unless you’re planning to work on something that will need to go fast on 1.0/1.1 HW (unlikely as 1.3-based Teslas are already announced), get a GTX260. The GTX280 is better but is more expensive and the power requirements may cause problems in some hosts.

Ah, I was under the impression that PCIe 1.0 cards would work in a 2.0 slot, but not vice versa (at least in some cases). Thanks for the info.

I appreciate the recommendations everybody! It seems that I can get the GTX260 for less than $300, which is great.

Can’t wait to run some of my research on it.

As double precision goes, does double precision run as fast as single precision? I’m not super up on my single vs double precision in processor pipelines.

Double precision is significantly slower than single. It’s advisable to only use it where necessary (utilize mixed precision computations).

Each multiprocessor has 8 single-precision units and 1 double precision unit. So double is about 8 time slower to calculate. But given the fact that kernels are mostly memory-bandwidth limited, you will mostly have a lower performance for doubles because of higher register usage. And if you want to use doubles as input or output, you suffer from a double amount of memory that needs to be transferred, so will get about half the performance.

GTX 260s are very cheap right now. $250 on Newegg, 230 after rebate. ATI’s really putting the screws to NVIDIA this generation. And both are getting sued for collusion, so they decided to cool it with the price fixing for a bit :haha: