AMBER Archive (2009)

Subject: [AMBER] MD on GPU: an existing commercial implementation

From: Sasha Buzko (obuzko_at_ucla.edu)
Date: Sat Apr 11 2009 - 19:47:46 CDT

Dear Amber developers,

as a little follow-up to our earlier discussion.. I came across a
commercial package that claims to run MD on GPUs and support multiple
simulation features along with reading Amber and Charmm input. It's also
being used in a distributed MD project gpugrid.net. The URL of the
commercial distributor: http://www.acellera.com/index.php?arg=acemd

Has anyone seen it and/or tried it for comparison with the current Amber
code? Can anyone comment on its accuracy?

Thanks in advance

Sasha

Ross Walker wrote:
> Hi Sasha,
>
>
>> Considering the emerging interest in GPU-based computational
>> applications, what do you think about efforts to port MD code to GPU
>> hardware? Are there any inherent limitations? Considering that accuracy
>> should be a higher priority than speed, are there any serious issues
>> with GPUs?
>>
>
> There are of course issues although I'm not sure that the scientific impact
> of these issues has been carefully studied yet so I'm not sure anyone
> actually definitively knows the answer to this. I would suggest that the
> biggest issue right now is a psychological one in the mind of some people
> implementing such things. With any accelerator or complex hardware there are
> always some parts of the problem that map better than others and right now
> the pressure to show that 'my code is faster than your code on a GPU' could
> lead some people to make algorithmic shortcuts that are somewhat
> questionable. Thus right now I would tread lightly when it comes to GPU
> implementations and demand internal consistency against 'trusted' cpu
> implementations.
>
>
>> Finally, if GPUs look like a reasonable platform to pursue, are there
>> any plans for of Amber code?
>> Since the advertised hardware acceleration sounds very tempting, it
>> would be interesting to know your opinion.
>>
>
> I have been discussing this with Vijay Pande and various people at NVIDIA
> and we are looking at ways of possibly hooking in their openMM GPU MD suite
> into AMBER. This is subject to the constraints that we can get it to support
> the same sort of functionality as PMEMD and more importantly pass all the
> test cases in the same fashion that PMEMD vs Sander does. Once there is
> something more concrete I'll let you know.
>
> All the best
> Ross
>
> /\
> \/
> |\oss Walker
>
> | Assistant Research Professor |
> | San Diego Supercomputer Center |
> | Tel: +1 858 822 0854 | EMail:- ross_at_rosswalker.co.uk |
> | http://www.rosswalker.co.uk | PGP Key available on request |
>
> Note: Electronic Mail is not secure, has no guarantee of delivery, may not
> be read every day, and should not be used for urgent or sensitive issues.
>
>
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER_at_ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
>
_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber