Re: PMF calculation using CPU and GPU

From: Souvik Sinha (souvik.sinha893_at_gmail.com)
Date: Fri Jun 08 2018 - 10:04:51 CDT

Yes, I see the doc and it is mentioned that the impact will be notable more
on scalar pressure than the energy and that is because of the atomic
contributions to the virial term as you already explained.

If I get you right, are you suggesting that the accuracy of the free energy
estimation on a GPU machine will depend on the choice of CVs? I mean, will
a simpler CV like distance is more likely to provide a more accurate
estimation using GPU (assuming CPU is always a better estimator in this
regard) ? I will definitely check the difference in the total force vs
total CV-force as you suggested.

Thanks for the reply.

On Fri, Jun 8, 2018 at 7:25 PM, Giacomo Fiorin <giacomo.fiorin_at_gmail.com>
wrote:

> First of all, check the doc of the new CUDA NAMD code to make sure you're
> up to date. With all due respect to the hard work put in to write it, NAMD
> 2.9 is nonetheless almost history in that regard.
>
> Changes in precision will affect averaging of thermodynamic quantities,
> but the magnitude of these effects depends on what you're averaging over.
> Averaging over terms with opposing signs (such as the atomic contributions
> to the virial pressure) is notably worse on single- and mixed-precision
> than in double-precision (pure CPU run). This is also mentioned at the doc
> page you link.
>
> For ABF, and in general every method where total forces are used for the
> TI free-energy estimator, you should try to figure out whether the total
> force on the collective variable (total CV-force) is a sum of very
> different terms. Comparing the variance of the total forces involved vs.
> the square of the total CV-force may be a way of doing it. If the two are
> different by many orders of magnitude, the accuracy will be lower with GPUs.
>
> Giacomo
>
> On Fri, Jun 8, 2018 at 9:36 AM, Souvik Sinha <souvik.sinha893_at_gmail.com>
> wrote:
>
>> Hi,
>> I was wondering, for converged PMFs (calculated using ABF-MW
>> simulations), will there be any difference in the landscapes calculated
>> with GPUs and without GPUs?
>>
>> In the thread 'http://www.ks.uiuc.edu/Research/namd/2.9/ug/node88.html',
>> it is mentioned that the force calculation slightly differs in GPU
>> accelerated runs. Does it really mean that convergence I attain using
>> CPU-only calculations (along a specific reaction coordinate) will not be
>> the same if I repeat it with some GPU resources? I guess it will but not
>> sure of the reasons.
>>
>> Any advice will be helpful. Thanks
>> --
>> Souvik Sinha
>> Research Fellow
>> Bioinformatics Centre
>> Bose Institute
>>
>> Contact: 033 25693275
>>
>
>
>
> --
> Giacomo Fiorin
> Associate Professor of Research, Temple University, Philadelphia, PA
> Contractor, National Institutes of Health, Bethesda, MD
> http://goo.gl/Q3TBQU
> https://github.com/giacomofiorin
>

-- 
Souvik Sinha
Research Fellow
Bioinformatics Centre (SGD LAB)
Bose Institute
Contact: 033 25693275

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2019 - 23:19:59 CST