Does copying from gpu to host generally take longer than copying from host to gpu?

Hi,

I am using the parallel computing toolbox, and it seems like copying data to a gpu using gpuArray() generally takes much longer than copying data back to the host using gather(). For example, if I try:

A = rand(500,500, 50); 

tic, 
B=gpuArray(A); 
toc, 

tic, 
gather(B); 
toc,

Then the gather() takes about 0.055 seconds while the gpuArray() takes only 0.018 seconds. Is this behavior expected? Am I using the wrong method to time this?

Thanks.

Answer 1

[Jill Reese]

I think you have arrived at the same conclusion as this blog post on GPU performance. It provides a lot of detail on how to properly benchmark GPU operations in MATLAB.

Loren's Blog on GPU Performance

[Adam]

Thanks for the tip! That blog is very helpful. Does anyone have any idea why the gpu to host copy should be so much slower?