GPU performance with short vectors

Hello - I see GPU computation underperforming when used for vector manipulation with short lengths.

>> a = rand(1000000, 100,'gpuArray'); 

>> b= gather(a);

>> tic; for i=1:100 ; eval('q = zeros(1000000,1);for i = 1:100; q = b(:,i)+q;end') ; end;doc

Elapsed time is 45.489811 seconds.

>>tic; for i=1:100 ; eval('qq = zeros(1000000,1);for i = 1:100; q = a(:,i)+q;end') ; end;toc

Elapsed time is 0.875140 seconds.

same when done for short vectors see GPU computation under performing:

>> a = rand(200, 100,'gpuArray');

>>b= gather(a);

>> tic; for i=1:100 ; eval('q = zeros(200,1);for i = 1:100; q = b(:,i)+q;end') ; end;doc

Elapsed time is 0.021727 seconds.

>>tic; for i=1:100 ; eval('qq = zeros(200,1);for i = 1:100; q = a(:,i)+q;end') ; end;toc

Elapsed time is 0.833865 seconds.

Any insight will be appreciated.

Thank you.

Answer 1

[Joss Knight]

Computation in a GPU core is significantly slower than in a modern CPU core. It makes up for that by having a lot of them - thousands. If you don't give it thousands of things to do at once, you're never going to beat the CPU.

In your simple computation above you are unnecessarily using a loop. This may have been for illustrative purposes, but if it reflects your actual code, you will gain back your performance by removing the loop, i.e.

q = sum(a, [], 2);

Answer 2

[Walter Roberson]

Do not use eval() for this. use timeit()

[Edric Ellis]

... and for GPU timings, use gputimeit.

[Walter Roberson]

Good point.

[MatlabNinja]

Thank you for your insight. time and gputimeit gives very similar results and shows similar trend where smaller vector(a & b above) had worse run performance when run on GPU.