Re: [AMBER] repeat amber run on cuda

From: Ross Walker <>
Date: Fri, 05 Jul 2013 13:58:07 -0700

Hi Kanin,

If you rerun a GPU run with identical settings, which includes the value
of ig the random number seed, then you should get bitwise identical
results. I.e. if you restart each simulation from the exact same restarts
etc. (if you are using a GTX-Titan then all bets are off since these are
not currently supported).

If you rerun with different values of ig then you will get different
results. In this case the difference in your results will reflect how well
converged your statistics are. So if you did indeed rerun but had ig set
to a different value this would explain why you see the differences you
do. It would also tend to suggest that you are a LONG way short of being

These days when reviewing papers containing MD simulations I tend not to
accept them unless they have a decent amount of work to properly estimate
how sampling errors affect their results. I typically suggest one always
runs a minimum of 5, ideally more, runs from different starting
structures, or conditions and uses the results of these to show proper
error bounds on the results. This is particularly important for things
like binding energy calculations which, to be honest, without enough
repetitions to properly estimate error bars are of very limited use.

Good luck,

All the best

On 7/5/13 1:35 PM, "Kanin Wichapong" <> wrote:

>Dear Amber Users,
>I have some questions about running MD on GPU. I finished the simulation
>and calculated BFE. But I lost some files so I did the simulation again
>using the same starting structure, same parameters and same computers/gpu.
>However, I didn't get the same results. Before I got MMPBSA around -50
>kcal/mol but now I got around -29 kcal/mol and it is similar to MMGBSA. I
>compared the structure derived from the last snapshot, but it is also not
>really similar.
>I know that repeat MD runs will not get the same exactly results, however,
>it should be somehow similar, especially when you use the same computer
>same MD parameters, right? So, I just wonder how can I believe in the
>results or something wrong.
>Best Regards,
>Kanin Wichapong
>AMBER mailing list

AMBER mailing list
Received on Fri Jul 05 2013 - 14:00:03 PDT
Custom Search