Double precision throughput on GTX's

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi I am looking to buy a good graphics card to do scientific computing with CUDA. Hence I will need a good double precision throughput for the card
for running my CUDA applications.

Tesla cards is too costly for me. So my final candidates for purchase (opinions welcome) are GTX 460, 560. I would like to know how good the double
precision performance AND support for these cards are in terms of :

–double to single precision performance ratio for these cards
–flop per second

Thank you

2

Double precision to single precision peak arithmetic throughput ratio is 1/12th for GTX 460 and GTX 560. It’s 1/8th for GTX 470, GTX 480, GTX 570 and GTX 580 which might thus be a better fit for you. You can find peak single precision GFLOP/s values on Nvidia’s web page or on Wikipedia.

3

I would strongly recommend not going with either the GTX 460 or the GTX 560. I’ve found the former to behave very strangely under CUDA (sometimes performing as much as 50% below expectations) and the latter is based on a similar architecture. Do yourself a favour and go with the GTX 570.