( Update: have posted a MUCH simpler method of driver install. Steps for CUDA toolkit install have to be followed as given in this post, i.e. , bulleted step # 10 – 19 )
Although the topic has been addressed succinctly in a CentOS forum post, there are certain things like plymouth configuration post Nvidia driver install, etc. which I felt needed to be jotted down for reference. So, here we go describing the Nvidia CUDA toolkit installation on a CentOS system:
- Download the appropriate toolkit, driver and SDK from Nvidia’s website.
- RHEL, and its derivatives come with the open source Nvidia driver called nouveau. Before installing Nvidia drivers, we need to ensure nouveau drivers dont get loaded. For this, append the following in the line starting with ‘kernel’ in the file /etc/boot/grub.conf:
- Install the Development Tools and Development Libraries group packages, and a few extra packages listed below:
sudo yum groupinstall ‘Development Tools’ ‘Development Libraries’
sudo yum install kernel-devel gcc-c++ freeglut freeglut-devel libX11-devel mesa-libGLU-devel libXmu-devel libXi-devel gcc* compat-gcc* compat-glibc* compat-lib*
- Restart the system. Upon restart, you’ll see that the resolution of the display would have gone for a toss. Thats due to blacklisting the nouveau driver, and is a sign that we are on track! Open terminal and type the following to goto non-GUI mode (called, runlevel 3):
sudo init 3
- Above command takes us to text mode. Change directory to /usr/src/kernels/ and note down the complete path of the kernel folder present. In our scenario, it shows up as:
- Change directory to the folder containing the downloaded files from Nvidia’s website (say ~/Downloads). Mark the 3 downloaded files as executables:
chmod a+x NV*; chmod a+x cuda*; chmod a+x gpu*
- Now finally, we are ready to run the installer. First is the Nvidia Driver install :
sudo sh NVIDIA-Linux-x86_64-295.20.run –kernel-source-path=/usr/src/kernels/2.6.32-220.13.1.el6.x86_64/
- NOTE : there’s a double minus sign before the word kernel above. During the above install, accept the licence agreement shown. Reboot upon completion:
- You’ll notice that the GUI resolution is back to normal, indicating successful Nvidia driver install. Now, cudatoolkit has to be installed.
- Open terminal and change directory to ~/Downloads. Run the cudatoolkit*.run file:
sudo sh cudatoolkit_4.0.17_linux_64_rhel6.0.run
- During the install, you’ll be asked to supply installation path. Enter the default path itself (/usr/local/cuda).
- Once completed, few more steps are needed, like adding /usr/local/cuda to default path environment variable, etc. :
sudo nano /etc/ld.so.conf.d/cuda.conf
- Add the following lines to the above created file :
- Save the above file by pressing Ctrl+x, followed by ‘y’ and pressing Enter. Now run:
- For adding cuda install path to enviroment path variable, edit ~/.bash_profile file using a text editor (say, nano ~/.bash_profile) :
export PATH=($PATH: /usr/local/cuda/bin)
export PATH=($PATH: /usr/local/cuda/lib)
- Finally, gpucomputingsdk needs to be installed. For that :
- During the install , you’ll be asked for install path. Keep in mind that the sdk can take around 400-500MB. Say, we install it to ~/Documents/NVIDIA_GPU_Computing_SDK.
- Once done, we need to compile the files in the SDK:
- To check whether everything is working fine, we’ll run the deviceQuery file, provided by the SDK just installed:
- You should see something similar to what is shown at this link .
Congrats for successfully installing the Nvidia driver and the cudatoolkit on your CentOS system. I know, things should be much simpler. I wish someday Nvidia open-sources their drivers and CUDA toolkit to make things simpler for Linux enthusiasts. All that is left now, is to fix the ugly white-blue scrolling bar that shows up instead of the beautiful Plymouth at boot.
- Reboot system. At the grub prompt, press ‘e’ against the first item on the grub menu to edit its kernel arguments (this mode is called kernel edit mode).
- Out the the three lines being shown (starting with : root; kernel; initrd ), scroll to kernel line and press ‘e’ again. Type in the following at the end of that line:
- Press Enter after typing the above and press ‘b’. You’ll see a tabulated list of keywords against several screen-resolutions. Note down the number being shown against the most appropriate screen resolution. Say, the number is 361. Now, reboot system. Again, enter the kernel edit mode described in the previous two list items. The only difference being that instead of vga=ask, enter vga=0×361. Now press ‘b’ and you’ll see the beautiful Plymouth back in its glory!
- To make this change permanent, open terminal and open /boot/grub/grub.conf as sudo (sudo nano /boot/grub/grub.conf).
- Find the line saying ‘kernel’ against your most recent kernel version and add vga=0×361. Save the file (Ctrl+x, y, Enter)
Plymouth theme will now show up everytime that you’ll reboot. Enjoy!