is F@H single or double precision?
Posted: Mon Apr 27, 2009 9:06 am
is F@H single or double precision?
Community driven support forum for Folding@home
https://foldingforum.org/
woohoo \o/John Naylor wrote:All cores are single precision except the three DGromacs cores which are double precision, as said by ^w^ing.
There's someone who know's GPUs. Here's another bit of trivia for you to tuck away...metal03326 wrote:I guess the GPU cores are single, because they run on HD 2x00, which do not have double precision.
mhouston wrote:It's all currently single precision. All 800 on the 48XX SPs can support single precision floating point, leading to a 2.5X increase in peak performance. Double precision on the 48XX is also 2.5X the performance of the 38XX.
VijayPande wrote:With a lot of effort, one can rewrite code that depends on double precision to run in single. It's rare that the whole code requires double -- usually there's just a few precision sensitive parts. By treating these parts carefully (e.g. reordering the numerics to avoid things like adding big numbers to small numbers, etc.) single precision can always work, but with a lot of effort in some cases.
We've already done that effort in GPU1, so the GPU2 code benefits from this today.
Indeed. Quite.alpha754293 wrote:Good to know. Thanks guys.