Can I install and use CUDA without installing NVIDIA graphics driver first? Thanks.
A simplified answer to this is “no”. The CUDA driver relies on low-level functionality that certain components of the graphics driver provide. Strictly speaking, that low-level functionality could also be seen as components shared by the graphics driver proper and the CUDA driver, but in practical terms these components are distributed as part of the graphics driver package.
What is the motivation for your question? What are you trying to achieve?
Thanks for the answer.
My desktop has standalone NVIDIA cards and motherboard-integrated Intel graphics. I want to use the Intel chip for display and NVIDIA cards only for computing. I guess I still need to install the NVIDIA graphcs driver in addition to CUDA then? BTW, I am using Debian.
I do not have experience with that scenario. On platforms with integrated Intel graphics I have always ignored if and, where necessary, used an older low-end NVIDIA card (such as a Quadro 600) as a dedicated graphics resources in addition to a high-end compute-only Tesla card.
I am under the impression that Linux allows multiple graphics drivers to be active. You may want to inquire on a Debian forum about this.
You can do this:
- leave the NVIDIA GPU out of the system, install your linux distro, and get everything running the way you want it. (The display will be using the intel graphics, obviously.)
- remove any traces of the nouveau driver from the system (as root):
echo -e “blacklist nouveau\noptions nouveau modeset=0” > /etc/modprobe.d/disable-nouveau.conf
- power down, install the nvidia GPU, power up
- Install the nvidia GPU linux driver using the runfile install method, selecting “no” when prompted to update the xorg.conf file
- Install the desired CUDA toolkit runfile installer, selecting “no” when prompted to install the nvidia driver.
The X display system should be using your intel graphics, and should not “know anything” about your NVIDIA GPU.
Are you sure cuda requires nvidia drivers? I installed cuda without the display drivers and seems working fine…
I am certain that it is required to have a proper GPU driver installed if you wish to run a CUDA code on that GPU. If you did something like:
sudo apt-get install cuda
that will also install a driver.
Yea, It had installed the driver. But how to exactly install nvidia drivers with ubuntu 1604. I had tried almost every thing but ending up in login loop every time. Is there any fail safe procedure?
Thanks in advance.
The login loop usually comes about in my experience when you have a graphics adapter that you are using besides an NVIDIA GPU, in addition to the NVIDIA GPU in your system. For example intel graphics.
The login loop is solvable and there are a number of questions here on this forum detailing how. You will find them if you search on login loop. For example, do a google search on “nvidia devtalk login loop” and take the first hit and start reading.
Thanks. I just moved on to Ubuntu 1404 and installed. Everything went on so smooth. I don’t know why there were many issues with 1604.
Any specific reason why Ubuntu 16.04 is stuck in login loop. I blacklisted the Nouveau drivers still in login loop for 16.04 where as it is successful for 14.04.