Best (Quality/Price) graphics card for heavy scientific computing

Hi All.

I need to buy a Graphics card which I will use mostly for heavy data processing (CUDA based) and also for high-resolution visualisation. With the current cards available, the Quadro 6000 (latest) seems very appropriate, but it’s very expensive. Any alternative? Thanks for your suggestions.

Regards

Khem

What kind of data processing are you doing? Do you need a lot of device memory? Do you need the fastest double precision performance?

Hi Siebert,

Thanks for your reply. I indeed need a lot of device memory and surely fast double precision performance. I am doing real-time processing of 1024*1024 (or highe resolution) images.

Thanks again

Khem

Quadro is not a CUDA-oriented line. That would be Tesla. NVIDIA charges a big premium for their “professional” lines (Quadro and Tesla). The premium is so huge that many (I’d say, most?) budget-conscious people use desktop cards instead.

Generally speaking, Tesla cards have two big advantages: fast double precision performance (which is crippled in regular desktop cards) and error-correcting memory. If you can live without these two, you can get a desktop GeForce card like GTX 470 or 560 that will match the Tesla in everything except double precision performance, for about one tenth the price. Or you can get a massive water-cooled dual 590 kit (5 single-precision teraflops, 700 watt peak power consumption; Tesla C2050 only makes 1 SP teraflop) and it’ll still be cheaper.

How much memory do you really need? 1024x1024 pixels is not a lot, unless you’re going to keep hundreds of them in the memory at the same time. Even the cheapest CUDA capable GPU’s commonly have 768 or 1024 MB of memory. Can you imagine a task where 1024 MB would not be enough?

Can you describe in a little more detail what kind of processing and visualization you will be doing?

Are you SURE you need double-precision for processing and visualization? I understand the need to try to have everything as accurate as possible, but if your output is an 8-bit image, it does not make sense to do FP64 calculations and then losing all the accuracy downsampling to 8-bit for output anyways.

Also, I work in image processing field as well (more specifically: bio-imaging analytics), but we run some pretty sophisticated segmentation, registration, and tracing algorithms. Unless you are analyzing thousands of those images in one go, you are better off putting your money into faster CPUs which will speed up everything else.

Tesla is Quadro without all the display output hardware. They cost roughly the same too, so for all intents and purposes, you may as well grab a Quadro.

Is this still true? The Tesla C1060 didn’t have a display connector on the back, but now the C2050/C2070 do.

Tesla C2070 only has one display connector, Quadro has three. And Quadro can work with SDI capture & output cards, for the professional video audience. On the other hand, Tesla is specced to eat 10% more power than Quadro, at the same clocks and amount of memory…