On GPUs, AMBER is limited to a finite range of precision. When doing initial minimization, there could be large forces / gradients that surpass the precision. This can lead to failure. We generally recommend minimizing on the CPUs. If that fails, try minimizing with electrostatics turned off first (on the CPUs). As the equilibration (minimization) step is such a small part of the workflow, why not run it on the CPUs (especially since minimization has never been optimized on the GPUs).
Reminder - single simulations are not statistically significant - plan on running multiple replicates with different initial conditions (3 at least) and show consistency.
--tec3
________________________________________
From: Anita Meena <Anita.Meena.cys20.chemistry.iitd.ac.in>
Sent: Wednesday, March 23, 2022 11:05 PM
To: amber.ambermd.org
Subject: [AMBER] [AMBER ] cudaMemcpy GpuBuffer::Download failed an illegal memory access was encountered.
Dear Amber users,
I am trying to simulate a protein in water. I have successfully
generated the topology (.top) and coordinate (.rst) files using tleap.
I am getting the following error while minimization, particularly in the
minimization of water.
cudaMemcpy GpuBuffer::Download failed an illegal memory access was
encountered.
And this didn't give results in the output file (min_wat.out) and didn't
generate the coordinate file.
I used the following command for minimization:
pmemd.cuda -O -i min_wat.in -o min_wat.out -p job_solv.top -c
job_solv.rst -r min_wat.rst -ref job_solv.rst.
----------------------------------------------------------------------------------------------------
Here is the min_wat.in, I used
Minimize water
System minimization:
&cntrl
imin=1, ntmin=1, nmropt=0, drms=0.1,
maxcyc=10000, ncyc=7000,
ntx=1, irest=0,
ntpr=500, ntwr=500, iwrap=0,
ntf=1, ntb=1, cut=5.0, nsnb=5,
igb=0,
ibelly=0, ntr=1,
restraintmask="!:WAT", restraint_wt=10.0,
&end
/
----------------------------------------------------------------------------------------------------
I also tried the same for 3 steps using pmemd.cuda (setting ntpr=ntwr=1)
but the same problem popped up.
When I did it using the sander.MPI and it worked.
Can someone explain, why it's showing problems with pmemd.cuda.
Thanks and regards,
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Mar 23 2022 - 22:30:03 PDT