I have only tried going completely to 32 bits with gromacs. My experience
there showed that if you use single precision floats and mpi, results become
nonreproducible almost immediately. Really, with the various operations
involved, and the sheer volume of them in a single step, this would be
expected. I was interested in this mostly from the perspective of cutting
communications volume, but because being able to validate software as
correct is rather important, I decided against this fairly quickly. So my
"take" on this is that while we have nowhere near the precision in our
calculations as what is represented in double (or even single) precision,
the rounding errors, especially in combination with networking indeterminacy
(or just getting the same results from two different platforms, two
different compilers, two different sets of compiler options, or two versions
of the s/w with logically equivalent but different mathematical operations)
makes using single precision floating point unattractive. I did try some of
this also in developing the precursor to pmemd; the cost in loss of
precision vs. benefit in speedup was not worth it, at least for the
platforms I tried it on.
Regards - Bob Duke
----- Original Message -----
From: "Cordova, Luis E." <cordoval.ornl.gov>
To: <amber.scripps.edu>
Sent: Thursday, June 09, 2005 1:39 AM
Subject: AMBER: precision switch from 64 to 32 bit (double from single
precision)
>I would like to pose the following question about precision to the AMBER
>community:
>
> What is the impact of changing precision from 64 bit to 32 bits in Amber?
>
> Thanks,
>
> \Luis
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Thu Jun 09 2005 - 13:53:00 PDT
