Re: [AMBER] gpu and cpu runs

From: Ross Walker <>
Date: Wed, 16 Jul 2014 10:03:59 -0700

Hi Mary,

To offer a little extra info here. There is no 'problem' with the random
number stream being different, it just means that the random friction
forces applied by the Langevin thermostat will be slightly different, this
leads to a divergence in the trajectories - it will not however affect any
statistical results. If you want to observe the effect that differences in
the random number stream have on a trajectory then you can just run two
CPU simulations (with ig=-1) and compare those. You will see the magnitude
of the differences should be equivalent to what you see for the GPU run
compared with the CPU run.

The take home message is that no single trajectory in its own right should
be considered to be definitive.

All the best

On 7/15/14, 8:29 PM, "Mary Varughese" <> wrote:

>thank you very much sir
>mary varughese
>On 7/16/14, Jason Swails <> wrote:
>> On Tue, Jul 15, 2014 at 6:57 PM, Mary Varughese <>
>> wrote:
>>> sir,
>>> above mentioned topic link:
>>> you have said ; if you are using NTT=3 the random
>>> number generator on the CPU is different from that used on the 3 GPU
>>> which would account for the differences in such short runs.
>>> i also used ntt=3 ; i feel there is much difference in my 15 ns cpu and
>>> gpu
>>> runs. That's why i asked ; which valid ntt value could be used which
>>> no such problem.
>> ​The problem is not your thermostat. The problem is that you expect
>> results to converge within 15 ns.
>> Even running NVE dynamics with good energy conservation, a GPU and CPU
>> simulation started from the same starting conditions will eventually
>> diverge because the system is chaotic and floating point round-off
>> are enough to trigger this divergence.
>> ​
>>> Also i would like to know if i can compare values (like free energy)
>>> simulation ; one run on sander and other run on pmemd(cuda)?
>> ​Yes.
>> HTH,
>> Jason
>> ​
>> --
>> Jason M. Swails
>> BioMaPS,
>> Rutgers University
>> Postdoctoral Researcher
>> _______________________________________________
>> AMBER mailing list
>AMBER mailing list

AMBER mailing list
Received on Wed Jul 16 2014 - 10:30:02 PDT
Custom Search