MathGroup Archive 2011

[Date Index] [Thread Index] [Author Index]

Search the Archive

Re: CUDA computing

  • To: mathgroup at
  • Subject: [mg121504] Re: CUDA computing
  • From: David Bailey <dave at>
  • Date: Sun, 18 Sep 2011 04:09:07 -0400 (EDT)
  • Delivered-to:
  • References: <j51sl8$pdq$>

On 17/09/2011 11:26, dan jack wrote:
> Do anyone use Nvidia GEForce GT520 or GT540 to perform parallel
> computation with CUDAlink?
> I search for a hint to buy one of these with 48 or 96 cores, but don't
> know how much improvement in calculations I will benefit.
> I'm involved with learning algorithms and neural netoworks so I think
> there is room for parallelization.
> What do you think about?
> And do you think it's better invest money on big CPU or big GPU or
> both?
> thank you all
In my opinion, the first step should be to ensure your code is 
efficient. The speed of Mathematica code can vary enormously with quite 
subtle changes.

For example, if the critical part of your NN calculation can be 
expressed as matrix multiplication, try to do that, and make sure you 
use packed arrays, or sparse arrays as appropriate.

David Bailey

  • Prev by Date: Re: Will it be a problem to include the mathematica plots in conference
  • Next by Date: Re: String Deletion
  • Previous by thread: CUDA computing
  • Next by thread: simple graphics question