Re: Alchemical calculations in CUDA version of NAMD?

From: Lela Vukovic (lvukov1_at_gmail.com)
Date: Wed Mar 17 2010 - 01:28:50 CDT

Hello,
originally I asked a question about alchemical free energy
perturbation calculations (TI) in CUDA built NAMD. They ran and the
ti.out file was written out. However, the ti.out output is different
from the output obtained from CPU calculations (every other output of
dG/dl (vdw) is 0.0000, and dG/dl (Coul) is quite different).
>From these results, I assume that this method has not yet been
implemented fully into the NAMD, as stated in the tutorial.
Does anybody know if this methodology is going to be implemented in
CUDA-built NAMD, and if yes, are there any time estimates?

Thank you,
Lela

> In the NAMD tutorial, it says that alchemical free energy perturbation
> calculations are not available in CUDA built NAMD.
> However, thermodynamic integration calculations do run in CUDA-built
> NAMD, and, in my experience, they are not accelerated as the
> calculations performed for regular NPT dynamics of the same system.
> Since energy evaluation is done on CPUs, and in these calculations,
> the output is the derivative of energy, I tried decreasing the
> frequency of output for this derivative. But, the calculation speed
> does not seem to be affected by this.
>
> So, my question is if CUDA-built NAMD is going to have alchemical
> calculations available to run at accelerated rates? If not, is there
> an intrinsic limitation to accelerating this type of calculations by
> use of GPUs (as only nonbonded force evaluation is performed on GPUs)?

2010/3/3 Lela Vukovic <lvukov1_at_gmail.com>:
> Hello again,
> the difference turned out to be in the PME settings. With same PME
> settings, the calculation speeds of TI calculations and regular
> dynamics in NPT ensemble are similar in CUDA built NAMD.
> Lela
>
> 2010/3/3 Lela Vukovic <lvukov1_at_gmail.com>:
>> Hello,
>> In the NAMD tutorial, it says that alchemical free energy perturbation
>> calculations are not available in CUDA built NAMD.
>> However, thermodynamic integration calculations do run in CUDA-built
>> NAMD, and, in my experience, they are not accelerated as the
>> calculations performed for regular NPT dynamics of the same system.
>> Since energy evaluation is done on CPUs, and in these calculations,
>> the output is the derivative of energy, I tried decreasing the
>> frequency of output for this derivative. But, the calculation speed
>> does not seem to be affected by this.
>>
>> So, my question is if CUDA-built NAMD is going to have alchemical
>> calculations available to run at accelerated rates? If not, is there
>> an intrinsic limitation to accelerating this type of calculations by
>> use of GPUs (as only nonbonded force evaluation is performed on GPUs)?
>>
>> Thank you,
>> Lela Vukovic
>> ~~~~~~~~~~~~~~~~~~~~~~~~~~
>> PhD candidate
>> Department of Chemistry
>> University of Illinois at Chicago
>> 845 W Taylor St., Room 4500
>> Chicago, IL 60607
>> Tel. (312) 996-1531
>> ~~~~~~~~~~~~~~~~~~~~~~~~~~
>>
>>
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 05:22:49 CST