I’ve experienced a bit of difficulties when I’ve tried to enable CUDA in my workstation. Those were mostly related to system lags while I’ve been performing CUDA computations. That was because Gnome/Xserver were using NVIDIA card. I’ve realised you’d be much better of using your discrete graphic card for the system and leaving NVIDIA GPU only for serious tasks 🙂 Note, this will disable NVIDIA GPU for GNOME / X11 and also for gaming, so be aware…
Below I’ll describe briefly how I’ve installed NVIDIA drivers and configured Ubuntu 18.04 with Gnome3 and Xserver for comfortable CUDA computations.
The best if you install CUDA toolking and drivers before you plug the card, as just plugging the card may cause issues with running Ubuntu otherwise (it did in my case). In order to install NVIDIA drivers, just follow official Nvidia guide.
Then after reboot plug the card to your computer and in the BIOS select integrated card as your main card. In my BIOS it was under
Advanced > Built-in Device Options > Select Boot card > CPU integrated or Nvidia GPU.
If you experience any problems, uncomment WaylandEnable=false in
/etc/gdm3/custom.conf to use X11 for GDM and Gnome. Don’t do that, if you plan to use Wayland!
Now make sure you have Nvidia plugged in and working.
# show available graphic cards
lspci -k | grep -A 2 -i "VGA"
If you installed the drivers from NVIDIA website, you may need to restore java
sudo rm /etc/alternatives/java jpath=/opt/java/jre1.8.0_211/bin sudo ln -s $jpath/java /etc/alternatives/java
Make sure to switch to integrated graphics card using either
- nvidia-settings > PRIME Profiles and select Intel (Power Saving Mode) (this should work for both, X11 and Wayland)
- or by editing
/etc/X11/xorg.confto something like that (if you use Wayland, this won’t work!)
Section "Device" Identifier "Intel" Driver "intel" Option "AccelMethod" "uxa" EndSection
Reboot your system and make sure Gnome isn’t using NVIDIA GPU (there should be no processes running on your GPU after reboot).
# check processed running on GPU nvidia-smi
Now, when you run any CUDA computation, your system shouldn’t be affected by high NVIDIA GPU usage.