Re: GeForce vs Tesla GPUs for high performance computing

From: Gianluca Interlandi (gianluca_at_u.washington.edu)
Date: Thu Apr 14 2011 - 14:48:58 CDT

I wonder which setup would be faster: a node with 4x 12-core opteron 2.3
GHz or a system with 4x GeForce GTX580 and a dual Quad core intel?

Something like this (unselect the one TESLA and purchase separately 4
GTX580):

http://www.siliconmechanics.com/i27103/Personal-Supercomputer.php

versus (select 12-core opteron 2.3 GHz):

http://www.siliconmechanics.com/i27727/1U-quad-opteron.php

Gianluca

On Thu, 14 Apr 2011, Axel Kohlmeyer wrote:

> On Thu, Apr 14, 2011 at 2:58 PM, Salvador H-V <chava09hv_at_gmail.com> wrote:
>> Hi All,
>>
>> If I understood well, to perform serious calculation for high performance
>> computing, people is using GPUs cards with double precision
>> such as tesla and quadro GPUs.
>
> that is not quite right. some methods and algorithms _have_ to have
> double precision math support, some others can make do with a combination
> where you do part of the computation in single, the rest in double and
> other calculations work just fine in single precision. the gromacs MD
> code has had an all-single precision option for ages and that is how you
> run it the fastest.
>
>> I was wondering if it is possible to use the gaming series of GPUs such as
>> GeForce GTX250 and the new one GeForce GTX580 for such kind of simulations.
>
> yes.
>
>> I mean, if i use this GTX series for small systems with NAMD how realiable
>> are the results?
>
> just as reliable as if you compute forces with SSE vectorization support.
> in the current implementation, you should not see a difference in the
> computation between Tesla and Geforce.
>
>> Are there any benchmark biomolecular simulation results between double  and
>> single precision GPUs?
>
> any g200 or fermi class GPU supports (some) double precision.
> the 20x0 tesla cards just have _more_ double precision units.
> that only makes a difference if _all_ calculations are done in double,
> but they usually are not, since even on Tesla, you have twice as
> many single precision units available than double precision.
>
> the differences of the tesla cards is:
> - more memory
> - ECC memory support (at a small performance penalty)
> - more double precision support
> - the cards are very tightly quality controlled
> and warranted by nvidia. if you have a broken
> one, you get it replaced instantly.
>
> all of this is very beneficial if you set up a large cluster
> with many gpus and a large variety of applications.
>
> if you want to equip a single desktop with a couple of
> high-end geforce GPUs and only want to use NAMD on it,
> then you might not need the tesla features so much.
>
> ECC support is a bit of a religious issue. you can have
> errors on the geforce without noticing, but many people
> use MD without a problem on them. if i was running
> MD on a geforce, i would regularly run cuda_memtest
> in order to confirm the card is still doing ok.
>
> cheers,
> axel.
>
>
>> Any help o suggestion (or point out to information), will be greatly
>> appreciated.
>>
>> Thanks a lot in advance,
>>
>> Salvador HV
>>
>
>
>
> --
> Dr. Axel Kohlmeyer
> akohlmey_at_gmail.com  http://goo.gl/1wk0
>
> Institute for Computational Molecular Science
> Temple University, Philadelphia PA, USA.
>
>

-----------------------------------------------------
Gianluca Interlandi, PhD gianluca_at_u.washington.edu
                     +1 (206) 685 4435
                     http://artemide.bioeng.washington.edu/

Postdoc at the Department of Bioengineering
at the University of Washington, Seattle WA U.S.A.
-----------------------------------------------------

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:20:08 CST