Sytem setup:
Ubuntu 14.04 64bit (fresh install)
Graphics card is an NVidia GeForce GTX Titan
The Titan being a relatively new card, it is a bit particular about its drivers. The Ubuntu kernel Nouveau driver obviously doesn't work for it, and you wouldn't be running CUDA with the kernel drivers anyway. I ended up sticking with the 331.62 NVidia driver version, because that is what comes with the CUDA installer. A newer version would probably have worked as well. The bottom line is, the Titan needs newer drivers. And, do yourself a favor and ditch Nouveau ;)
What didn't work
I broke the GUI trying various things that didn't work, so I'll talk about them here, to hopefully save you the trouble.
First off, when you try running nvcc, the CUDA compiler, you'll be told that you don't have it and you need to install it by running sudo apt-get install cuda-toolkit to get it. Don't do that! When you then try to run your cuda code and check for errors, you will get "No cuda-capable device detected". And you'll have no idea why, because you have a Titan card and that surely has to be cuda-capable! It is. You just don't have cuda-driver compatibility.
Another thing you shouldn't do is run sudo apt-get install nvidia-current. This and having the cuda-toolkit installed is what breaks everything.
What does work
This is how to do it right, step by step.
0. Purge all existing nvidia drivers, just in case.
sudo apt-get remove --purge nvidia*
sudo apt-get clean
1. Install build-essential if you don't already have it.
sudo apt-get update
sudo apt-get install build-essential
2. Download the cuda6 installer and unpack it.
I went with version 6.0.37, which can be found here:
http://developer.download.nvidia.com/compute/cuda/6_0/rel/installers/cuda_6.0.37_linux_64.run
Then, unpack it:
mkdir ~/Downloads/nvidia_installers;
cd ~/Downloads
./cuda_6.0.37_linux_64.run -extract=~/Download/nvidia_installers;
3. Switch to terminal mode, turn off the X server, and install the graphics driver
Ctrl+Alt+F1 to switch to the terminal.
cd ~/Downloads/nvidia_installers;
sudo service lightdm stop
sudo killall Xorg
sudo ./NVIDIA-Linux-x86_64-331.62.run
The driver installer will turn off and blacklist Nouveau and should then install without issues. Once installation is complete, you can turn on the driver and return to the graphical interface to install CUDA.
sudo modprobe nvidia
sudo service lightdm restart
4. Install CUDA and the Samples, then make sure everything works
sudo ./cuda-linux64-rel-6.0.37-18176142.run
sudo ./cuda-samples-linux-6.0.37-18176142.run
To check if this works, run the deviceQuery sample:
cd /usr/local/cuda/samples
cd 1_Utilities/deviceQuery
make
./deviceQuery
5. Add the CUDA binaries to your PATH
This tripped me up. When I went back and tried to compile my example code, I was again told that I don't have nvcc installed. What gives? Well, you just need to add it to the PATH. Simple.
To do this, you'll have to edit .bashrc. I also added a line to ./bash_profile to make sure ./bashrc actually gets used as the bash profile.
cd ~
vi .bashrc
//add the line: export PATH=/usr/local/cuda-6.0/bin:$PATH
vi .bash_profile
//add the line: source ./bashrc
_________________________________________________________________________________
Now that I was finally good to go, I computed the cubes of 64 numbers (0-64) - in parallel... on the GPU! On a device capable of and designed for handling thousands of threads. Yes, it was silly, because "the GPU doesn't even get out of bed in the morning for less than 500 threads!" But it proved that I can compile and run cuda code now, so it made me happy :)
Source: http://askubuntu.com/questions/451672/installing-and-testing-cuda-in-ubuntu-14-04
