Folding Forum

Would the use of double precision arithmetic actually improve the scientific return of any given F@H work unit - or is the science return currently limited more by inherent deficiencies or simplifications in the underlying physics model, rather than numeric precision?

Are there plans to use double precision arithmertics in future GPU client iterations, say GPU V3 ?

Christian

For the CPU, there are two versions of Gromacs, one single precisions and one double precision. DGromacs has been used in maybe 5% of the projects, so there must be some benefit because they won't use it unless it's needed. For the same protein, double precision runs slower since SSE can do 4 simultaneous floating operations and SSE2 can do two simultaneous doubles. Dr. Pande did answer a question about this once and he said they had revised the single precision code to obtain most of the benefits of double precision without the slower speed.

I don't know whether that applies to the GPU or not, but it might.

codysluder wrote:For the CPU, there are two versions of Gromacs, one single precisions and one double precision. DGromacs has been used in maybe 5% of the projects, so there must be some benefit because they won't use it unless it's needed.

What is the Gromacs 33 core, then? I am using it on one protein. It is a CPU core as well. Thanks!

rbpeake wrote:What is the Gromacs 33 core, then? I am using it on one protein.

http://fahwiki.net/index.php/Cores#Gromacs_33_core