VMD-L Mailing List

From: John Stone (johns_at_ks.uiuc.edu)
Date: Fri Nov 11 2011 - 12:41:03 CST

Hi,
  I thought I'd revisit the GPU memory capacity topic again now that
I've got the new "QuickSurf" representation running well and have more
experience with using it on large structures. This advice applies not
only to the new QuickSurf graphical representation in VMD 1.9.1, but
also to many of the MD trajectory analysis features that will be forthcoming
in subsequent VMD versions.

Besides the memory consumed for the atom coordinates associated with a
given frame, the new QuickSurf representation and many of upcoming
the analysis commands involve calculations that require building and
traversing uniform grids either for calculation of spatially localized
properties (e.g. densities, occupancies, electrostatics, orbitals, etc),
or for performing various neibhbor-neibhbor interactions and/or distance
tests. The grids have the potential to consume a lot of memory very
quickly. In the case of the QuickSurf representation, GPU memory capacity
is one of the primary factors that impacts performance. In cases where
the GPU doesn't have enough on-board memory to hold the entire model
and various intermediate density maps and triangle meshes, we have to
compute the surface in a piecewise manner, which causes additional
host-GPU data copies, and generally starts to slow things down.

For people interested in working on multi-million all-atom
molecular complexes, I would suggest getting GPUs that have
3GB of RAM if possible. There are various gaming-oriented
GPUs available that have 3GB of memory, so it is possible to get
GPUs with that much memory without going all the way up to the
"professional" cards that cost much more. I have an Asus laptop
with a GeForce 560M that has 3GB of memory, for example. I can
run most of the same calculations on the GPU on my laptop as I
can on a desktop workstation, mainly because I have sufficient
GPU memory capacity to do it. I would therefore suggest that anyone
that's looking to get good performance for visualization of large
structures make sure to buy GPUs with large memory capacity going
forward.

In my experience, a GPU with 3GB is a good compromise between price
and performance for working with large structures, and I would
expect such a GPU to remain usable for another 2 or 3 years given
what I know about the GPU computing and rendering algorithms I have
planned for upcoming versions of VMD.

Cheers,
  John Stone
  vmd_at_ks.uiuc.edu

On Wed, Oct 05, 2011 at 02:28:45PM -0500, Thomas Bishop wrote:
> Thanks JOhn,
> that's excellent info.
>
> Does this handy chart more or less work or is it to abusable?
> (usign 1024 as Gb conversion)
> Guess if the card only has 1Gb ram then cut the numbers by 3 or 4 for
> safety
>
> Atoms bytes GB Frames GB Ram Req'd
> 1 12.00 0.00 100000 1.14
> 1000 12000.00 0.01 100 1.14
> 10000 120000.00 0.11 10 1.14
> 100000 1200000.00 1.14 1 1.14
>
> Hi,
>
> On Sat, Oct 01, 2011 at 05:37:43PM +0400, Dmitry Osolodkin wrote:
>
> On 10/01/2011 04:58 PM, Axel Kohlmeyer wrote:
>
> you want to do simulations of 20,000,000 atom systems
> on a single workstation?
>
> Obviously not :)
>
>
> the first thing you have to worry about for visualization
> is not the GPU, but having sufficient RAM.
>
> This is a very important point. I'm working on adding the
> ability for VMD work with trajectory data entirely out-of-core, but
> in the short term until these features are completed, you will need
> enough system RAM to work with the portions of your simultion trajectory
> data that you need to view/analyze.
>
>
> And does the GPU RAM matter? I'm planning to start with Intel i7 CPU and
> 16 Gb RAM, and I have spare $1K for additional memory and GPU. What else
> should I take in mind when selecting GPU hardware? My aim is smooth
> interaction with large cartoon rendered system, its rotation etc.
>
> If you want to use the new GPU accelerated features of VMD,
> then yes, the amount of GPU memory you have will be very important.
> All of the new MD trajectory analysis features of VMD require
> enough GPU RAM to hold the equivalent of a few timesteps worth of
> coordinates. If you're working with 20M atom systems,
> then you need at least 1GB of unused GPU memory to do much of anything,
> but that memory also gets used by the windowing system, by
> recent versions of web browsers like Chrome, and so on. For things
> like time-averaging of coordinates, you might need 20 trajectory frames
> in memory at once. I have a big batch of new GPU analysis features
> coming in the next version of VMD this fall, and they will need
> GPUs with a decent amount of memory.
>
> You can of course use VMD without any of the GPU features, but if
> you want the GPU accelerated features, you'll need some GPU memory
> matched to the size of the structures you're working with.
> At least 12 bytes per atom, per timestep, plus various overhead...
>
> Cheers,
> John Stone
> vmd_at_ks.uiuc.edu
>
>
>
> a trajectory
> for systems of this size can become very large and as soon
> as your workstation will begin to swap, everything will
> become very slow.
>
> Yes, I know. The trajectory will be preprocessed before visualisation.
>
>
> the choice between nvidia or amd/ati depends
> a bit on the operating system. if you are using linux
> (which i would recommended) then the support for nvidia
> GPU is significantly better (using the nvidia provided
> drivers).
>
> Yes, I'm using OpenSuse linux.
>
>
> high-performance quadro cards are very expensive,
> but most of the features that they offer over geforce
> cards are not exploited by VMD.
>
> Just to make clear: which specific features of quadro cards are
> exploited by VMD? It's sort of curiosity.
>
>
> so you are probably best
> off with a GeForce GTX 580 card.
>
> Looks OK.
>
>
> cheers,
> axel.
>
>
> Thanks in advance and best regards,
> Dmitry Osolodkin
>
>
>
> --
> Dmitry Osolodkin, PhD
> Researcher
> Group of Computational Molecular Design
> Department of Chemistry
> Moscow State University
> Moscow 119991 Russia
> e-mail: dmitry_o_at_qsar.chem.msu.ru
> Phone: +7-495-9393557
> Fax: +7-495-9390290
>
> --
> *******************************
> Thomas C. Bishop
> Tel: 318-257-5209
> Fax: 318-257-3823
> http://dna.engr.latech.edu
> ******************************** 

-- 
NIH Resource for Macromolecular Modeling and Bioinformatics
Beckman Institute for Advanced Science and Technology
University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801
http://www.ks.uiuc.edu/~johns/           Phone: 217-244-3349
http://www.ks.uiuc.edu/Research/vmd/       Fax: 217-244-6078