Why GPU performance is worse than CPU for this code?

clear all
close all

tic
t=9999;

X = rand( t, 'single' );
    G = gpuArray( X );
    isequal( gather( G ), X )  % returns true
    classUnderlying( G )       % returns 'single'
    G2 = G .* G   
toc

   tic
  X = rand( t, 'single' );
      G1 = X;
      G3 = G1 .* G1  ; 
  toc

The execution time of GPU is less than CPu i5. Would anybody answer it?

[Jan]

About "clear all" see: http://www.mathworks.com/matlabcentral/answers/16484-good-programming-practice#answer_22301

Please decide, if the "GPU performance is worse than CPU", or if "the execution time of GPU is less than CPU".

Answer 1

[Walter Roberson]

There is a lot of overhead for sending data to a GPU. It is only faster for sufficiently big matrices.

Also, in the first one you are including the "isequal" and the "classUnderlying" in the timing, neither of which is needed for the computation. It is not fair to time different sets of activities.

[King Fisher]

You are right, I was unfair :)

Answer 2

[Jan]

Please try this:

X = rand(1e3, 1e3, 'single' );
G = gpuArray( X );
G2 = G;
tic
for i = 1:100
  G2 = G2 .* G;
end
toc

tic
G = X;
G2 = X;
for i = 1:100
  G2 = G2 .* G;
end
toc

[King Fisher]

thanks Jan !

Answer 3

[cdarling]

there is a possibility that your GPU supports parallel.gpu.GPUArray.rand(), and it would be much faster if you construct random numbers within GPU, instead of generate and copy them into GPU