Well... what a discussion :]
First of all: thank You for all Your comments.
They are really helpful.

I am aware of many interesting libs out there like lapack, blas, cgal, gmp, leda, mpfr and so on...
but the thing is that very often You just get to the point where You're stuck on the bug you canno't repair yourself.
As my older Colleagues suggested it is much better to use slower algorithms implemented by Yourself than highly optimized
code You cannot understand at all. If I implement everything myself it won't be so fast as the stuff I mentioned.
But I will be able to repair my code myself :] I won't have to wait (up to year) for libs next release until bug gets fixed for me.
Therefore I decided to buy a little cluster (8 GPUs 4 CPUs) to make it faster, instead of using "black box" code that I cannot control.

When huge commercial app is created (f.x. nastran, fluent etc) every part is made by a speciallist in the matter, under computer specialists control.
And every bug can be fixed without any problems.
I have to create quite huge thing myself so I want to keep as much control as possible. I will possibly include GMP, CGAL LAPACK in my app, but I also want
to have my own - not-super mega highly optimized efficient piece of code just in case, and to have choice if sth goes wrong

Thanks again,
I rated You

If your using GPUs for arbitrary arithmetic and need absolute reliability. I've heard people had issues with low level bit errors within such systems so that might be a concern. Here's a link to a presentation discussing it.


ECC for GPUs


-ddn

I think that you need to be more clear about what you consider to be a "bug" in this context.