espressomd-users
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [ESPResSo-users] Memory and compute requirements for LB GPU


From: Dominic Roehm
Subject: Re: [ESPResSo-users] Memory and compute requirements for LB GPU
Date: Wed, 09 Apr 2014 10:17:30 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:24.0) Gecko/20100101 Thunderbird/24.3.0

Hi Vincent,

since the LB grid is 3D the memory consumption increases with the power
of 3. A disadvantage of the current implementation is that is has to use
double buffering, which means that the entire grid has to be stored
twice, besides some other things like the forces and particle data....
(but the memory consumption due to the particles is negligible compared
to the LB mesh). With the 3 GB VRAM of your M2050 you figured out
225**3. Due to the cubic scaling you would end up with a 283**3 mesh
(grid space 1) on a GPU with 6GB and 357**3 with 12 GB. I don't know
what grid size or resolution you have in mind, but if want to use large
grids you need to use a cluster and the CPU code. Basically, a multi-gpu
implementation should be available in a few months, but since the code
got a lot of more features like Shan-Chen model and Electrokinetics it
takes some time.

Dominic


On 04/09/2014 01:52 AM, Vincent Ustach wrote:
> Hello all,
>
> I am working with large lattice Boltzmann systems. In the spirit of
> shopping for new graphics cards, it would be interesting to know how
> the LB GPU requirements on a graphics card scale with system size.
> Axel Arnold was able to answer a similar but more general question on
> August 9, 2013 so I thought I might ask.
>
> I think for each MD particle there are three linear interpolations,
> giving me some hope the computations for each particle scale with order N.
>
> In terms of memory, these days my simulation crashes above 225 X 225 X
> 225 nodes thanks to memory limitations on the Tesla M2050 I am
> currently using. At that size I am mostly just trying to crash the
> system :)
>
> I have the 2012 GPU paper for reference, as well.
>
> Best Regards,
>
> --Vincent Ustach
>   University of California, Davis


-- 
   Dominic Roehm
   Dipl.-Phys.
   Institute for Computational Physics
   Universit├Ąt Stuttgart
   Allmandring 3
   70569 Stuttgart, Germany 

   Office: 1.077
   Phone: +49(0)711/685-67705




reply via email to

[Prev in Thread] Current Thread [Next in Thread]