Re: [AMBER] cuda

From: Ross Walker <ross.rosswalker.co.uk>
Date: Mon, 24 Sep 2012 09:08:06 -0700

Hi Mary,

Unfortunately the nature of GPUs is such that one has to rewrite the
entire code to run on the GPU in order to obtain efficiency. Why we would
love to rewrite sander to run it would be a monumental task and we simply
don't have the resources or manpower to do it. One can do 'tricks' in
which just part of the calculation runs on the GPU (say the direct space)
but this needs stupidly big atom counts to see any real performance
improvement and also tends to be inefficient in terms of the number of
GPUs used. We have decided not to go down this route since we believe it
better to focus on accelerating the types of calculations that most people
run rather than going after simple headline numbers.

For this reason we used PMEMD as our base point since the code is much
cleaner and it was easy to implement the majority of the features without
many clashes. For example, sander has so many options we'd be forever
chasing which ones work with GPUs which don't etc and we would never
gotten to having a fully working code. Our intention is to slowly migrate
over features that are widely used (like TI for example) into PMEMD. This
is a much efficient use of resources than trying to hack the support into
Sander. Thus please stay tuned for more additions to PMEMD to come.

With regards to other parts of AMBER on GPUs. At present the code is just
in PMEMD. We are working on a library that we plan to release open source
under AmberTools for people to add GPU support to their own codes. Of
course this will suffer from the fact that copying back and forth from GPU
will destroy performance and so if one wanted to add their own specific
features this would likely need them to edit the cuda code to achieve
this. Unfortunately there are no magic bullets. So ultimately once the
library is completed the cuda code will be released under an open source
license. It will likely take us a while to get this done however.

All the best
Ross



On 9/24/12 7:41 AM, "Mary Varughese" <maryvj1985.gmail.com> wrote:

>Sir,
>
>Then sander would have no improvement in efficiency?
>Also the amber11 licenced is using or is there any other version for using
>with cuda.
>
>Thanking you
>On Mon, Sep 24, 2012 at 11:08 AM, filip fratev
><filipfratev.yahoo.com>wrote:
>
>> Hi,
>> Only pmemd.cuda.
>>
>> All the best,
>>
>>
>>
>> ________________________________
>> From: Mary Varughese <maryvj1985.gmail.com>
>> To: AMBER Mailing List <amber.ambermd.org>
>> Sent: Monday, September 24, 2012 7:03 AM
>> Subject: [AMBER] cuda
>>
>> Sir,
>>
>> Does installing AMBER11 with NVIDA CUDA on GPU machine works fine for
>> sander too for parallel. On going through details i notice only
>>pmemd.cuda
>> and not any sander.cuda.
>> Would you please clear this. It will help me to understand the
>>situation.
>>
>> Thanking you
>> --
>> Mary Varughese
>> Research Scholar
>> School of Pure and Applied Physics
>> Mahatma Gandhi University
>> India
>> _______________________________________________
>> 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
>>
>
>
>
>--
>Mary Varughese
>Research Scholar
>School of Pure and Applied Physics
>Mahatma Gandhi University
>India
>_______________________________________________
>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 Mon Sep 24 2012 - 09:30:03 PDT
Custom Search