Experience with setting up GPU for Keras and Tensorflow + Rants

11 Feb 2018

I’m going to document my experience with setting up Tensorflow on my GPU. I had hoped that I could setup a cloud version to see the differences in performance but unforunately the free tier of Google Cloud Compute won’t give me a GPU =(.

Basically, what it boiled down to is this:

  1. Had to update Ubuntu from 14.04 to 16.04 for the latest software from Nvidia.
  2. Download all the dependencies following this link[https://www.tensorflow.org/install/install-linux)

It seems pretty dumb to me that it’s such a painful process to set this up. You have to install the following dependencies:

  • Nvidia drivers
  • cuda
  • cuDNN
  • … and a bunch of other stuff I can’t remember.

Then you have to set the environmental paths. Remember, if something breaks along the process, as usual Google on.

So after all this trouble to use my GPUs for learning purposes, this is the final result:

(Insert pic here from other computer)

The good part: a 4x speed of in neural net training with two 970s!

The bad part: HAVING TOO MANY TABS IN CHROME AND RUNNING THE GPU ACTUALLY CAUSES COMPUTER TO FREEZE…