GPU for ML only on Ubuntu

I fought with NVIDIA drivers taking over my graphics and xterm on Ubuntu 20.04. It was either have my GPU do everything and lose up to half its memory in the process… or nothing. I had it working once… then updates uninstalled the driver and broke everything again.

That was a few weeks ago. I’m confident now, a few updates later, that I’m using the correct driver and that settings are all good, so I’ll document what it took. Some may be unnecessary steps but since I got it working once and haven’t experimented further, I can’t say.

• in BIOS settings, I had to enable iGPU multi-monitor and enable CPUgraphics

• install CUDA tools

  sudo apt install nvidia-cuda-toolkit

• install cudnn from NVIDIA; requires an account

• install a headless driver and accompanying utils

  sudo apt install nvidia-headless-440 nvidia-utils-440

• add to PATH and LD_LIBRARY_PATH in .bashrc

  export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
  export LD_LIBRARY_PATH=/usr/local/cuda/lib64

• create a new config file in /etc/X11/xorg.conf with settings

  Section "ServerLayout"
      Identifier     "Layout0"
      Screen      0  "Screen0"
  EndSection
  
  Section "Screen"
      Identifier     "Screen0"
      Device         "intel"
  EndSection
  
  Section "Device"
      Identifier      "intel"
      Driver          "intel"
      VendorName      "Intel Corporation"
      BusId           "PCI:0:2:0"
      Option "AccelMethod" "sna"
      Option "TearFree" "true"
      Option "DRI" "3"
  EndSection

• modify grub settings in /etc/default/grub so that

  GRUB_CMDLINE_LINUX="nogpumanager"

  then run sudo update-grub for the settings to take effect

• if, in the process, the file /usr/share/X11/xorg.conf.d/11-nvidia-prime.conf was created: delete it

• finally, last but not least, following random/good internet advice, add (create?) the following to /etc/modprobe.d/blacklist-nvidia.conf

  blacklist nvidia-drm
  alias nvidia-drm off

Now all should be working. To be sure, running nvidia-smi should show something like

Fri Sep 11 19:45:45 2020       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.66       Driver Version: 450.66       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce RTX 2070    Off  | 00000000:02:00.0 Off |                  N/A |
| 35%   47C    P0    22W / 175W |      0MiB /  7982MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

And, in e.g. a jupyter notebook, running tf.config.list_physical_devices('GPU') should output the

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Eureka! All the memory for ML!