CPU computation time vs. batch size, almost linear with slope=1 ?

35 views

Skip to first unread message

unread,

May 19, 2016, 9:50:55 AM5/19/16

to Caffe Users

I'm using caffe with caffenet model and 12Gib GPU.

To my surprise the computation time vs. batch size graph is almost linear with slope=1.

The computation is solver.step(10) (10 iterations of solver), using default solver

I have expected both computation time to have a much lower slope

100 images take 16 sec
300 images take 44 sec
600 images take 80 sec
2000 images take 260 sec

more notes:

- In a second test I tested forward pass only. same linear results

- I already made sure caffe is utilizing the GPU ( by using nvidia-smi command )

Any advice/idea for what may I be wrong with?

Isn't my expectation for a much lower slope correct?

unread,

May 19, 2016, 12:51:38 PM5/19/16

to Caffe Users

Hi,

Check your Blast/Openblast etc' optimization,

בתאריך יום חמישי, 19 במאי 2016 בשעה 15:50:55 UTC+2, מאת Doron Shuldiner:

Reply all

Reply to author

Forward

0 new messages