I tried to install many times via the .run file, but some error always crept in and I either ran into a login loop or completely lost the display. Therefore I would recommend to use the .deb file and not fiddle with the display manager.

NVIDIA CUDA Installation Guide for Linux
is an excellent link that lists the complete details.Make sure you follow each step as it is given .

To install the Nvidia driver you can do the following:

1. In Ubuntu "Search your Computer" menu at the left top corner search "Additional Drivers" (You might also do System Settings->Software and Updates->Additional Drivers)




2. In the menu that appears select one of the Nvidia Drivers and click "Apply Changes".(This step uses the internet.If it still fails then your proxy server might be blocking the download)




3. Reboot your system.




4. Open a terminal window and type nvidia-smi. If your driver has been installed correctly you should see something like :

+------------------------------------------------------+                       

| NVIDIA-SMI 3.295.41   Driver Version: 295.41         |                       

|-------------------------------+----------------------+----------------------+

| Nb.  Name                     | Bus Id        Disp.  | Volatile ECC SB / DB |

| Fan   Temp   Power Usage /Cap | Memory Usage         | GPU Util. Compute M. |

|===============================+======================+======================|

| 0.  Tesla C2050               | 0000:05:00.0  On     |         0          0 |

|  30%   62 C  P0    N/A /  N/A |   3%   70MB / 2687MB |   44%     Default    |

|-------------------------------+----------------------+----------------------|

| Compute processes:                                               GPU Memory |

|  GPU  PID     Process name                                       Usage      |

|=============================================================================|

|  0.  7336     ./align                                                 61MB  |

+-----------------------------------------------------------------------------+

You can easily install CUDA according to the previous link now. In brief:

sudo apt-get install linux-headers-$(uname -r)

Download a toolkit from here and then install the .deb file (replace name accordingly)

sudo dpkg -i cuda-repo-<distro>_<version>_<architecture>.deb

then run:

sudo apt-get update

sudo apt-get install cuda

I also tried difference approaches so as to install Cuda 8.0 in Ubuntu 16.04. Finally, these are the steps which do the trick. I followed this tutorial and updated corrected steps as follows.

1. 
   

   Update the system

   
   
   

   apt-get update && apt-get upgrade 
   
   

   
   


2. 
   

   Download VirtualGL and install it. To install

   
   
   

   dpkg -i virtualgl*.deb
   
   

   
   


3. 
   

   Download and install CUDA 8.0　and install it. I suggest to do it vs through the internet. As like this,

   
   
   

   　

   
   


4. 
   

   Install required dependencies.

   
   
   

   apt-get install linux-headers-$(uname -r)
   
   apt-get install freeglut3-dev libxmu-dev libpcap-dev
   
   

   
   


5. 
   

   Update system PATH in .bashrc which can be found in the home directory. Please note if you install those thing into difference location, please update path according to that.

   
   
   

   export PATH=$PATH:/opt/VirtualGL/bin
   
   export PATH=$PATH:/usr/local/cuda/bin
   
   

   
   


6. 
   

   Install bumblebee-nvidia and primus.

   
   
   

   apt-get install bumblebee-nvidia primus
   
   

   
   


7. 
   

   Edit the bumblebee config file so bumblebee knows we are using the NVIDIA driver. Please update the path according to your system. Here is reference view which will help.

   
   
   

   sudo nano +22 /etc/bumblebee/bumblebee.conf 
   
   

   
   
   

   Add:

   
   
   

   [bumblebeed]
   
   ServerGroup=bumblebee
   
   TurnCardOffAtExit=false
   
   NoEcoModeOverride=false
   
   Driver=nvidia
   
   XorgConfDir=/etc/bumblebee/xorg.conf.d
   
   Bridge=auto
   
   PrimusLibraryPath=/usr/lib/x86_64-linux-gnu/primus:/usr/lib/i386-linux-gnu/primus
   
   AllowFallbackToIGC=false
   
   Driver=nvidia
   
   [driver-nvidia]
   
   KernelDriver=nvidia
   
   PMMethod=auto
   
   LibraryPath=/usr/lib/nvidia-367:/usr/lib32/nvidia-367
   
   XorgModulePath=/usr/lib/xorg,/usr/lib/xorg/modules
   
   XorgConfFile=/etc/bumblebee/xorg.conf.nvidia
   
   Driver=nouveau
   
   [driver-nouveau]
   
   KernelDriver=nouveau
   
   PMMethod=auto
   
   XorgConfFile=/etc/bumblebee/xorg.conf.nouveau
   
   

   
   


8. 
   

   Run the following and record the PCI address of your video card.

   
   
   

   $ lspci | egrep 'VGA|3D'
   
   00:02.0 VGA compatible controller: Intel Corporation Device 5916 (rev 02)
   
   01:00.0 3D controller: NVIDIA Corporation Device 179c (rev a2)
   
   

   
   


9. 
   

   Edit the xorg.conf.nvidia file so it knows the PCI address(01:00.0 for me) of your video card. Update PIC address as below under section "ServerLayout"

   
   
   

   sudo nano /etc/bumblebee/xorg.conf.nvidia
   
   

   
   
   

   Add:

   
   
   

   Section "ServerLayout"
   
       Identifier  "Layout0"
   
       Option      "AutoAddDevices" "false"
   
       Option      "AutoAddGPU" "false"
   
       BusID "PCI:01:00.0"
   
   

   
   


10. 
    

    Reboot the system and have a fun with running some sample codes.

    
    
    

    sudo shutdown -r now
    
    

    
    

This is a looooong answer as i was screwing my laptop several times while writing it. However, I rather to keep it long since it maybe useful for other people too ;)
The best part of my answer starts from Edited-Updated

Sooooo, I read all the answers here and other places, I dont know why, but each of them cause me an issue :(

After 4 days, re installing Linux back and forth here is the way that worked for me.

Before going to the main procedure i want to mention an alternative method.

alternative method if you are using a laptop:

So you can switch between your nvidia and your intel gpu on your laptop by using

sudo prime-select intel

sudo prime-select nvidia

In other words, you can switch to intel and install nvidia and the switch back to intel for normal usages and whenever you want to use deep learning switch to nvidia one.

Anyway,

let me talk about the main method that finally works for me (info here are mainly grabbed from Link):

Deleting and purging all existing nvidia/cuda stuff:

sudo apt-get remove --purge nvidia-*

sudo apt-get purge nvidia-cuda*

sudo apt-get purge nvidia-*

sudo /usr/bin/nvidia-uninstall

sudo /usr/local/cuda-8.0/bin/uninstall_cuda_8.0.pl

sudo rm -rf  /etc/modprobe.d/blacklist-nouveau.conf

Then, we just update everything:

sudo apt-get update

sudo apt-get upgrade

sudo apt-get dist-upgrade

sudo reboot

Now, there would be a possibility that you cannot log in and you get stuck in the loop...

No worries, I faced that more than 50 times...

press ctr+alt+F2
type your username and password

now type these:

sudo service lightdm stop

Optional, some people also need to type this, honestly idk what is the use of it: sudo init 3

sudo nano  /etc/modprobe.d/blacklist-nouveau.conf

add

blacklist nouveau

options nouveau modeset=0

to it and save it and exit.

Then type:

sudo update-initramfs -u

go to the file that you have cuda .run file there and type:

sudo sh cuda_8.0_linux.run --override

sudo service lightdm start

sudo reboot

Sooo, if you are lucky, you should be able to login now. As you may guess, i was not a lucky one, and I still could not f**** login.
So i had to press ctr+Alt+F2 again and do the following:

sudo ubuntu-drivers autoinstall

sudo reboot

Now i could login finally.

Now it is the time to set the paths and check the installations.

type:

export PATH=/usr/local/cuda-8.0/bin${PATH:+:${PATH}}

export LD_LIBRARY_PATH=/usr/local/cuda-8.0/lib64                         ${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}



nvidia-smi

nvcc -V

it should show you that you have cuda 8.

Also just in case you can also do these:

export PATH=$PATH:/usr/local/cuda-8.0/bin

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-8.0/lib64





gedit ~/.bashrc

add these at the end:

export PATH=$PATH:/usr/local/cuda-8.0/bin

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-8.0/lib64



# Added by me on 2013/06/24

PATH=~/bin:$PATH

export PATH

Save and exit gedit.
Type:

sudo ldconfig /usr/local/cuda-8.0/lib64

A question for the people who knows more than me:

So everything sounds like working but when I shutdown/reboot my system for a few seconds it shows

So everything sounds like working but when I shutdown/reboot my system for a few seconds it shows “Failed to start Load Kernel Modules” I tryied this post but it is not helping, please let me know if you know how to fix it.

----------------

Edited-Updated

Guess what, I screwed again.

But this time I came with a much easier solusion. and here is the main point: Sometimes we need to say NO

Here is what works really good for me. after you purge and remove everything and sudo reboot do this:

sudo ubuntu-drivers autoinstall

sudo reboot

press ctr+alt+F2
type your username and password

now type these:

sudo service lightdm stop



sudo sh cuda_8.0_linux.run  

IMPORTANT: during the installation, the first question asks for reinstalling the driver again, SAY NO to this mother f**** question
Say yes to the rest of them though :D
after finishing.

sudo service lightdm start

press `alt+ctr+F7` 

login to your dear PC

Did it work? Your Welcome :)

I've written a blog post on this a while ago - Nvidia CUDA toolkit installation - ubuntu 16.04 LTS - notes/

My environment: Dual boot Windows 10 and Unbuntu 16.04 LTS.

Copying and pasting here some major learnings. Please see blog post for detail instructions (just to avoid duplicate)

Major learning:

• Disable Secure Boot at UEFI Firmware setting (do this in BIOS mode / restart from Windows advanced startup). (I did try for ages hoping I could get it working with Secure Boot enabled. No luck Secure Boot stopped the Nvidia driver from installing properly on Ubuntu. Disabling Secure Boot turns out to be the only working solution for me - if you are able to get Nvidia driver installed without having to disable Secure Boot, do let me know)


• Follow the Linux CUDA Installation Guide.


• (opinionated...) Use the Linux .deb (package manager) installation (for simplicity.). Download the .deb file in a browser. Install from terminal command line.


• (opinionated...) Don't use runfile installation (too complicated).

 Detail Instructions:

Please refer to Nvidia CUDA toolkit installation - ubuntu 16.04 LTS - notes/

The steps that worked for me:

1. sudo apt-get install nvidia-cuda-toolkit OR 1'. the installation instructions here




2. You need to select from Software&Updates/Additional Drivers the nvidia driver (375, in my context)




3. Follow the blue screen when you restart and only from there disable secure boot by inputting your password set during nvidia driver install. (any secure boot disabling through the BIOS did not work for me).

Now the installation test output is successful.

I got it to work after reading several posts:
I had an ATI card in the computer already which turned out to be very useful. I installed GTX 1070 along side of the ATI and started installing Kubuntu 16.04. Only the display connected to the ATI card had image initially, which allowed me to install the driver NVIDIA-Linux-x86_64-367.27.run downloaded from the vendor's website. To install CUDA, I downloaded the cuda_7.5.18_linux.run file. I installed the cuda toolkit by using two switches:

cuda_7.5.18_linux.run --silent --toolkit

The cuda samples can also be installed from the .run file. One issue was cuda does not like gcc5. So I did sudo apt-get install gcc-4.8 and then changed the default gcc to this version by:

cd /usr/bin/

sudo unlink gcc

sudo ln -s gcc4.8 gcc

sudo unlink g++

sudo ln -s g++-4.8 g++

I replaced gcc to gcc5 after cuda is installed. Compiling the cuda samples also need to be done with gcc4.8, gcc4.9 might work but I did not try it.

A generally preferred method is to install SW is via deb files when available as they provide a more robust way to handle dependencies and a more reliable method for removing SW. The CUDA 8.0 release-candidate was available for 16.04 (in the dev zone) that way and now the CUDA 8.0 for Ubuntu 16.04 is available via deb files (local) and (network) :https://developer.nvidia.com/cuda-downloads

Just a kind reminder, Ubuntu 16.04 might not install cuda at the assumed location /usr/local/cuda-8.0.61. Hence export PATH=/usr/local/cuda-8.0.61/bin${PATH:+:${PATH}} may not work.

When I was trying to install "cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb" on Ubuntu 16.04, I simply followed the instructions here http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#post-installation-actions. However, I was not able to compile
cuda-install-samples-8.0.61.sh home
or nvcc -V

It turned out that Ubuntu installed cuda in /usr/local/cuda-8.0 instead of the assumed location /usr/local/cuda-8.0.61. Hence I changed export PATH=/usr/local/cuda-8.0.61/bin${PATH:+:${PATH}} into export PATH=/usr/local/cuda-8.0/bin${PATH:+:${PATH}} and I successfully installed cuda.

The accepted answer didn't work for my case. I was installing CUDA 8.0 on my labtop with following specifications:

• Graphics Card: GeForce GTX 950M (cc 5.0)


• CPU: Intel Core i7-6700HQ (with Intel HD Graphics 530)

The following guide installs the NVIDIA driver first, and then installs CUDA 8.0.

Installing CUDA 8.0 on a fresh installation of Ubuntu 16.04

1. Launch [Software & Updates]. Select [Additional Drivers] tab.


2. 
   

   In the list, find your graphic card. Among the drivers that can be used for the card, choose the proprietary driver from NVIDIA. Then press [Apply Changes] button. In my case, under the graphics card name "NVIDIA Corporation: GM107M [Geforce GTX 950M]", there were two selections:

   
   
   
   
   
   • Using NVIDIA binary driver - version 375.66 from nvidia-375 (proprietary, tested)
   
   
   • Using X.Org X server - Nouveau display driver from xserver-xorg-video-nouveau (open source)
   
   
   
   



3. Delete default installed video drivers with $ sudo apt remove xserver-xorg-video*.



4. Reboot.


5. Download CUDA 8.0 Toolkit from here. Among the installer types, choose "runfile (local)". This downloads cuda_8.0.61_375.26_linux.run.


6. Run the installer with $ sudo sh cuda_8.0.61_375.26_linux.run.
   
   
   
   
   • [Optional] If your currently installed NVIDIA driver version is higher than the driver version contained in the downloaded installer, you can choose not to install the driver while installing CUDA. In my case, since I already have driver version 375.66, which is higher than 375.26 contained in the installer, I chose not to install.
   
   
   
   


7. After install, config your binary path and library path (You can follow the directions from the instller). If you choose to configure ld.so.conf and the following error occurs: libEGL.so.1 is not a symbolic link, follow the direction from this link.

I initially tried doing that sudo lightdm stop stuff, but it lead to a login loop. So I found a new method:

1. Copy the file cuda_9.0.176_384.81_linux.run (in my case it was runfile) to any directory in /home/<your_username> like Downloads or Documents or anywhere.




2. After that restart your computer and when Ubuntu boot menu appears go to 'Advanced Options → Recovery Mode' (if it does not appear hold down shift key while booting)



3. 
   

   Select 'drop to root shell', press ENTER to proceed when asked for pressing enter or Ctrl-D.

   
   
   

   Edit: Run mount -o rw,remount / to get read-write priviliges.

   
   



4. Go into that directory where you have copied the cuda installation file.




5. Run the command on the basis of type of file, it can be found at https://developer.nvidia.com/cuda-downloads after selecting your desired target as you have done earlier. In my case it was sudo sh cuda_*.run




6. This is important step and proceed slowly and carefully, when the long information/agreement ends ACCEPT it.




7. Then it will ask about the NVIDIA DRIVER INSTALLATION press yes(y).




8. Then it will probably ask about OpenGL libraries installation, skip it because it may override your normal driver installation and cause problems, in my case it did. So Press no(n).




9. Then go ahead with all the installations and it will complete automatically and at last show the logfile in /tmp.




10. Now reboot the system by entering the reboot command at the recovery mode shell.



11. 
    

    After your system starts it might not show the CUDA sample files, because you need to complete these two mandatory post-installation steps :

    
    
    

    
    

    [A] Add the correct path for cuda.

    
    
    

    [B] Add correct path for LD_LIBRARY_PATH

    
    

    
    
    

    Add the path to ~/.bashrc file and run source ~/.bashrc to make the path permanent so that it didn't vanish after reboot, confirm it by closing the current terminal and running the second command in step 12 again in another terminal.

    
    
    

    Refer to Go to 7. Post-Installation Actions

    
    


12. 
    

    To check whether CUDA is installed properly or not run both of the below mentioned commands and check if nvcc -V give output or not

    
    
    

    cat /proc/driver/nvidia/version
    
    
    
    nvcc -V
    
    

    
    


13. 
    

    Go to ~/NVIDIA_CUDA-9.0_Samples/1_Utilities/deviceQuery, then run these:

    
    
    

    make
    
    
    
    ./deviceQuery
    
    

    
    
    

    and match the output with this Image, your might be different but the output format should match.

    
    



14. Congrats you installed CUDA Toolkit successfully. After that go here and try some examples Go to 7.2 Recommended Actions .

COURTESY - CUDA TOOLKIT DOCS

P.S - Any type of criticism is welcome, apologizes in advance for any mistakes, this is my first answer on askubuntu.com.

THANK YOU SO MUCH FOR READING:)

This worked for me

sudo rm /tmp/.X*-lock

sudo apt-get purge nvidia-*

sudo reboot

sudo service lightdm stop

Press Alt + f1

sudo rmmod nvidia

sudo sh cuda_8.0.61_375.26_linux.run 

sudo service lightdm start

and reboot

Having done this multiple times, successfully/unsuccessfully loosing my display, coming here - gaining insights - some cuda libs not in path, missing , not installed - the sane way is to just install the linux drivers for your nvidia-card https://medium.com/techlogs/install-the-right-nvidia-driver-for-cuda-in-ubuntu-2d9ade437dec
and work on nvidia-cuda docker images - base or devel.

Do volume mapping from your code folder to the container - install what you want -
Same with working with keras or tensorflow or just pure opencv

docker run --net=host --runtime=nvidia -it -v ~/coding:/coding nvidia/cuda: /bin/bash

Note TF also comes with its docker