timer Result of matrix multiplication

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,
I’ve installed cuda (toolkit and SDK) and visual studio 2005 and i want to test examples.
I open project matrix multiplication with visual studio, compile it and run it OK without any problem.
But I found that the time of a matrix multiplication is not the same for the same matrix. For example with 3*3 matrix i have many result (0.11ms, 0.10 ms …)
I want to know if any one have the same problem and what is the solution
Thanks

2

3*3 matrix is so small, only one trhead block is used.

try large dimension, for example, 1024 * 1024 or larger

3

Thanks for reply
With large dimension matrix i don’t have the same result timer for many run ( for example matrix 512*512 the difference is only 1 ms (15829.870111 ms for the first run and 15828.432129 ms for the second).
As you see the difference is so small but I just want to know if the variation of result timer of the same code don’t make a problem to implement my own program and I risk to abtain erroneous result.
Thanks

4

Sorry I obtain this result (time) when i used 1 block but if i use 16*16 the time is shorter but the time is not the same for many run!!!

5

if you use all the resources (all SMs), then you can measure your kernel in average sense,

then variation should be less than 1%.

However if you under-utilize the GPU, then time variation is large.