My platform is D870, our program need double precision supporting. But D870 does not support double precision. Although CUDA 2.0 support double precision, it simply handle double precision as single single. right? How to handle this kind of problems. Thanks!
You can emulate double precision using two single precision floats. Check the Mandelbrot example.
Well, it depends when you need double precision. If it is for adding up big & small numbers, you can use Kahan’s. I also am very suspicious when people say they need double precision. I have thought the same a few times, and always was able to rewrite my algorithm to not run into issues.
Otherwise there is no other option than buying a GT200-based card.