Cuda on Titan X cannot work well on Ubuntu

Accelerated Computing CUDA CUDA Programming and Performance
1

I bought a Titan X recently. But it seems that it doesn’t work well. Here are some symptoms:

  1. When I call the gpuDevice in matlab, it needs a long time to show the information. While when I do the same thing on the same computer with Titan X changed to K40, the information are shown immediately.

  2. When I call the paralleldemo_gpu_backslash from matlab, it works well on small size matrix. But when the size of the matrix increased to be 9216 * 9216, the whole system, the OS will not response to anything. The mouse won’t move, the shortcuts cannot be used. It seems the whole system breaks down.

The specific setting of my system is as follows:
ubuntu 14.04 12.04
CUDA 7.0
Drivers I tried: 346.47 346.72 346.59

Sorry I am not experienced with these issues and I really appreciate any help and advice.
Thank you.

2

the first thing that comes to mind is that both matlab and the os are independent variables in your case and observation; you have chosen to focus on the os, and not matlab…

you might want to consider experiments and test cases that remove the 2nd variable, to only have 1 variable at a time, to really get to the root cause
other tests not using matlab on ubuntu
using matlab with another os other than ubuntu

3

To 2.)
If the Titan X is the display board and you run heavy compute tasks on that, your graphics performance will be affected.
The Tesla K40 is a dedicated compute board and not driving the display. Running compute tasks on that alone will not affect the display performance handled by another GPU.

4

I think matlab variables are of type double by default.

K40 will have significantly better performance than Titan X when processing double quantities, in general.