Build TensorFlow 1.2 from source with CUDA 8.0 and Cudnn 6.0 on Ubuntu 16.04

I want to compare the difference between native pip install and build from source for TensorFlow in terms of computation speed.

Setup

Basically, I first followed this blog post:

Install GPU TensorFlow From Sources w/ Ubuntu 16.04 and Cuda 8.0

However, 2 additional issues to solve:

1. Need to use Bazel 0.5.2 instead of latest version. Latest version failed to build during my time.
   1. sudo apt-get purge bazel
   2. go to this link for installation of 0.5.2.
2. If facing issue of “ImportError: libcuda.so.1: cannot open shared object file: No such file or directory“, will need to follow below steps:
   1. sudo apt install nvidia-361-dev
   2. sudo find /usr/ -name 'libcuda.so.1' (then you will know path of libcuda.so.1) If it’s already inside /usr/local/cuda dir, no need for the 3rd step.
   3. just copy the libcuda.so.1 to /usr/local/cuda/lib/
   4. Most importantly, reboot your machine!

If everything goes well, you can test your gpu like in this link.

Experiment

I did a comparison between two VMs on Google Cloud – one using ‘pip install’ and another using build from source – with same hardware config and everything else. I ran a basic CNN model on CIFAR10 image set, and when using GPU, there is no difference in computation time. However, when I switched off GPU and used only CPU for computations, I noticed a reduced of processing time around 25% – 30%.