Nvidia driver, CUDA tools and libraries

nvidia-settings
Oh no, another Nvidia driver repository? Why?

This driver reflects my personal view for the way the driver should be packaged for Fedora and CentOS/RHEL. It’s somewhat different from ELRepo repositories for RHEL/CentOS and from RPMFusion packages for Fedora.

Repository installation

To install the repository on a supported Fedora distribution, run as root the following command:

dnf config-manager --add-repo=https://negativo17.org/repos/fedora-nvidia.repo

To install the repository on CentOS/RHEL:

yum-config-manager --add-repo=https://negativo17.org/repos/epel-nvidia.repo

Starting from Fedora 25, the Nvidia software packages are available for installation by default also in Gnome Software.

Please note that the driver will show up only if your system matches one of the PCI ID supported by the driver. Otherwise, only the other Nvidia programs (mostly for CUDA development) will show up in the software center.

What’s different?

First of all the packaging is a lot simplified; more stuff is compiled from source, smaller packages and more options. This packages try to comply as maximum to the Fedora Packaging Guidelines; which means they have debuginfo packages, default Fedora’s GCC compile time options (where possible) and standard locations for binaries, data and docs.

What follows below, is a detailed explanation of all the “differences” from the various Nvidia driver packages that I was able to spot on the web and a detailed description on how to install components, etc.

Nvidia drivers

Packaging

  • nvidia-settings, nvidia-persistenced, nvidia-xconfig and nvidia-modprobe are compiled from source.
  • All RPM filters except for GL and OpenCL libraries have been removed, so there is no weird dependency option in the SPEC file. RPM pulls in all correct requirements on its own. This is to avoid pulling in the Nvidia drivers instead of the Mesa libraries or in place of the new open source OpenCL support that’s in Fedora.
  • Simplified packaging with much simpler and readable SPEC file.
  • Dependency on libva-vdpau-driver. So in Totem, or any other libVA supported application you can benefit from VDPAU acceleration.
  • Sources are generated with a script and inserted individually in the various packages; so it can be easily reproduced just by changing the version and rerunning the script.
  • nvidia-xconfig is not required on anything that uses the modular X.org directives, as it writes too much in the configuration file (keyboards, monitors, etc.) and the required entries should be written in separate configuration files under /etc/X11/xorg.conf.d. The package is still available as it’s required to speed up some configuration like multi-monitor setups with SLI Mosaic enabled from the command line, but not installed by default.
  • The NVIDIA OpenGL-based Framebuffer Capture (NvFBCOpenGL) libraries (NvFBC and NvIFR) are private APIs that are only available to NVIDIA approved partners for use in remote graphics scenarios (i.e. Steam In-Home Streaming hardware encoding); so they are packaged in another small package called nvidia-driver-NvFBCOpenGL.
  • The nvidia-settings package now builds the external libXNVCtrl.so library that can be used to control the graphic cards through the NV-CONTROL extension. This library updates the old and obsolete one in Fedora based on drivers version 165.
  • Starting from version 343.13, the nvidia-settings binary is compiled with GTK3 instead of GTK2 on Fedora and RHEL/CentOS 7+.
  • The driver can be installed separately from the nvidia-settings utility, so if you simply want a working driver and do not care about details, your experience should be as close as possible to the one with open source drivers.

Versioning

  • ELRepo ships 32 bit compatibility libraries in a separate package with x86_64 as the architecture and “32bit” in the name. 32 bit libraries should be like in RPMFusion, with an i686 package installable in parallel with the x86_64 one. There are no other packages in the distribution that are built for x86_64, with “32bit” in their name that contain i686 binaries (!), so Nvidia drivers should not be an exception. So no separate “32bit.x86_64” package for 32 bit libraries also on CentOS/RHEL; just install nvidia-driver-libs.i686.
  • Versions are not hidden; all packages have the same driver version.
  • No alternatives system, only the latest version which integrates CUDA support is available. For older releases nouveau works great; and anything below a GeForce 8xxx it’s in my opinion too low end to play anything modern. And Quake 3 and Doom 3 work greatly with nouveau, so that’s not a case!
  • The CentOS/RHEL repository contains the “Long Lived Branch version” where less changes occur; while Fedora repositories contains the “Short Lived Branch version”. Beta CentOS/RHEL and Fedora’s rawhide repositories will contain the “Beta Branch version”

CUDA support

  • CUDA libraries/tools for the driver are split into subpackages. There’s no need to install all the CUDA libraries and tools on a system that has only one adapter and is used for occasional gaming or for simple office use. This can save ~120 MB worth of installed libraries. nvidia-persistenced falls in this category as it’s not needed on a normal laptop or gaming system.
  • Complete packaged CUDA stack has been added for all supported distributions, all the packages provide/require/obsolete the relevant packages in the Nvidia CUDA repository; so you can enable this repository along with the official Nvidia CUDA one (x86_64 systems only).

Kernel modules

  • Multiple choice of kernel module packages; akmod (RPMFusion) for Fedora and binary kmod (Kernel ABI whitelists) for CentOS/RHEL. In addition to this, on both distributions dkms packages are available. This way all cases and personal preferences are covered for both distributions.
  • Starting from Nvidia driver version 334.16, the Nvidia DDX driver for X can also rely on the nvidia-modprobe command in the system to create devices and set permissions, so the new optional package has been added.
  • The nvidia module has a soft dependency on the nvidia-uvm module, making sure the module is loaded when installing the nvidia-driver-cuda package, but making sure that these modules are not included in the initrd (thing that would happen with systemd configuration (module-s-load.d). UDev rules make sure the module has proper permissions.
  • On Fedora, the kernel modules are compressed with XZ, like all the other kernel modules.

Default configuration

  • Dracut options are depending on the distribution; so no more “vga=normal is an obsolete option” at boot. Each distribution gets its own specific GRUB options for booting.
  • 96 DPI is written in the default xorg.conf config file. Why? Gnome 3 by defaults hard-codes a 96×96 DPI resolution, most of the free drivers do (intel, nouveau, etc.) as the EDID is almost never reliable (please see the excellent Adam’s Jackson post where he explains this). As an example, if you install the Nvidia drivers on a RHEL/CentOS 6 laptop where you used to have nouveau installed (96 DPI hardcoded), the fonts gets 90% of the time supersize and ugly as Gnome 2 and the Nvidia driver do not hard-code 96 DPI like Gnome 3.
  • Make X.org NVIDIA Files section to be loaded latest in case there are other packages providing a custom Files section.
  • Starting from Fedora 21, all driver X.org configuration can be managed by simply adding/removing X.org configuration snippets in /etc/X11/xorg.conf.d.
  • Use new OutputClass directive on Fedora 21 X.org server 1.16 (and later) to load the driver and do not rely on an edited /etc/X11/xorg.conf file. This also removes editing of the xorg.conf file from the package scriptlets. This does not hardcode the 96 DPI resolution.
  • Add the IgnoreABI directive by default on Fedora rawhide builds.

Kernel modesetting and Wayland support

Kernel mode setting on the nvidia-drm module has been disabled by default for various reasons. First of all, Wayland support in the drivers require a patched Wayland which has been refused upstream, and then the driver itself does not expose an FB driver for the console, so you won’t see any difference in the terminal output, you will still be limited to VGA.

There is a proposal for sorting everything out at XDC 2016 for hardware vendors that require to expose extensions for the drivers.

The Wayland libraries are still included in the Fedora builds, as all the dependencies are there but they are not used. On CentOS/RHEL 7 packages, they are not included as this would result in missing dependencies.

Vulkan support

vulkan_500px_june16Vulkan is now part of Fedora, so on supported Fedora releases, the Vulkan loader and libraries can be installed and you do not need to do anything to enable support in the drivers. CentOS and Red Hat Enterprise Linux do not have Vulkan yet. I’m not sure if it’s worth installing it by default along with the drivers, though.

Distribution and Nvidia driver version support

Here is a rundown of Nvidia supported drivers and options split by distribution. Basically, CentOS/RHEL will always get a Long Lived branch release if possible, Fedora always a Short Lived branch release, and unreleased distributions will always get a Beta driver.

Operating systemel6 / el7f27 / f28f29
Driver branchLong LivedShort Lived
Long Lived
Short Lived
Long Lived
Beta
Driver version390.77396.24396.24
NVENC8.0.148.0.148.0.14
Architectures:

x86_64
YesYesYes
Basic nvidia driver:

nvidia-driver
nvidia-driver-libs
nvidia-libXNVCtrl
YesYesYes
CUDA libraries and tools:

nvidia-driver-cuda
nvidia-driver-cuda-libs
nvidia-driver-NVML
nvidia-persistenced
YesYesYes
OpenGL Framebuffer Capture:

nvidia-driver-NvFBCOpenGL
YesYesYes
Nvidia tools:

nvidia-modprobe
nvidia-settings
nvidia-xconfig

YesYesYes
Binary kernel
modules (kABI):

kmod-nvidia
YesNoNo
DKMS kernel
modules:

dkms-nvidia
YesYesYes
aKMOD kernel
modules:

akmod-nvidia
NoYesYes
32 bit compatibility on x86_64:

nvidia-libXNVCtrl
nvidia-driver-libs
nvidia-driver-cuda-libs
nvidia-driver-NVML
YesYesYes
Development

nvidia-driver-devel
nvenc
nvenc-samples
YesYesYes
GLVND librariesYesYesYes
VDPAU libraries1.1.11.1.11.1.1

Optimus laptops

The driver should install and operate cleanly whether you are installing it on a system which has one or more discrete Nvidia cards or an Optimus laptop with an Intel and a Nvidia card. Nothing to do to enable or configure Optimus.

This is up to the point that when the drivers are installed, you can even turn off Optimus on or off in your system Bios (if your laptop allows that) and the only difference you should see is that there’s an additional VGA card enabled in your system (check with lspci) and that the Nvidia control panel switches between a PRIME Display, like in this picture:

And a normal RandR managed one, like in this one:

Everything else should not be different from your normal experience.

Limitations with the Nvidia driver

The limitations are the same as provided by the Nvidia driver, this means that if you are running it on an Optimus laptop, the Intel card can never power off. Which means higher power consumption, unfortunately. If you have an Optimus laptop and absolutely need the proprietary drivers, my suggestion is still to disable Optimus in the Bios.

Limitations with the OSS stack

On the contrary, if you use the OSS stack (nouveau/intel) the second card can be powered off if there’s no application running on it or display directly connected to one of the card’s outputs. That’s the best reason to use the OSS drivers at all if you you’re not doing serious gaming or 3D work:

$ sudo cat /sys/kernel/debug/vgaswitcheroo/switch
0:IGD:+:Pwr:0000:00:02.0
1:DIS: :DynOff:0000:01:00.0

You also got the nifty selection menu about running your game on the discrete card on Gnome, which is really cool:

It will power up the video card just before launching the process. Launching a program through that menu entry is like starting it from the command line with the DRI_PRIME variable declared. For example, the same as above would be:

$ DRI_PRIME=1 quake3 &
$ sudo cat /sys/kernel/debug/vgaswitcheroo/switch
0:IGD:+:Pwr:0000:00:02.0
1:DIS: :DynPwr:0000:01:00.0

As you can see, the discrete video card is turned on. For Steam, you still need to edit each of your game to run on the Nvidia card:

SLI systems

SLI is now enabled by default with the Auto profile, there’s nothing to do if you have a SLI system. If you need any different SLI option (AA, SFR, etc.), just override it in X.org configuration files.

Nouveau fallback

With the new expanded OutputClass support for X, as carried out by Hans, it’s now super easy to switch to the OSS stack if the proprietary Nvidia driver somehow does not work. No user space component is touched, as soon as the Nvidia kernel module is not loaded (check on /sys/module/nvidia), the desktop starts with the normal OSS components you get with a normal installation. Thanks to all the work done on libglvnd, the libraries loaded are the correct one for the driver you are running.

This means that the performance of the Nvidia card would be abysmal, but still you would get a nice desktop and browser to Google around for answers on how to fix it :).

Sample installation

Here is an example. Let’s assume you have a freshly installed Fedora 25 system with a recent Nvidia GPU and you want to:

  • Install the driver for gaming
  • Play Vulkan enabled games
  • Want to be comfortable with the control panel
  • Play 32 bit games on a 64 bit system
  • Play 32 bit Vulkan games on a 64 bit system
$ sudo dnf install nvidia-settings kernel-devel dkms-nvidia vulkan.i686 nvidia-driver-libs.i686
Last metadata expiration check: 0:33:49 ago on Mon Oct 24 14:14:30 2016.
Dependencies resolved.
============================================================================================
 Package              Arch     Version                                Repository       Size
============================================================================================
Installing:
 dkms-nvidia          x86_64   2:375.10-1.fc25                        fedora-nvidia   6.4 M
 libglvnd             i686     1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia   103 k
 libglvnd             x86_64   1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia   105 k
 libglvnd-gles        i686     1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia    29 k
 libglvnd-gles        x86_64   1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia    28 k
 libglvnd-glx         i686     1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia   114 k
 libglvnd-glx         x86_64   1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia   110 k
 libglvnd-opengl      i686     1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia    39 k
 libglvnd-opengl      x86_64   1:0.2.999-3.20161017gita813b56.fc25    fedora-nvidia    38 k
 libva-vdpau-driver   x86_64   0.7.4-14.fc24                          fedora           61 k
 libvdpau             i686     1.1.1-3.fc24                           fedora           35 k
 nvidia-driver        x86_64   2:375.10-1.fc25                        fedora-nvidia   3.1 M
 nvidia-driver-NVML   x86_64   2:375.10-1.fc25                        fedora-nvidia   397 k
 nvidia-driver-libs   i686     2:375.10-1.fc25                        fedora-nvidia    15 M
 nvidia-driver-libs   x86_64   2:375.10-1.fc25                        fedora-nvidia    14 M
 nvidia-libXNVCtrl    x86_64   2:375.10-1.fc25                        fedora-nvidia    26 k
 nvidia-settings      x86_64   2:375.10-1.fc25                        fedora-nvidia   935 k
 vulkan               i686     1.0.30.0-1.fc25                        fedora          1.5 M
 vulkan               x86_64   1.0.30.0-1.fc25                        fedora          1.4 M
 vulkan-filesystem    noarch   1.0.30.0-1.fc25                        fedora          8.0 k
 
Transaction Summary
============================================================================================
Install  20 Packages
 
Total download size: 43 M
Installed size: 184 M
Is this ok [y/N]:

As you can see, this system has dkms enabled kernel module and libraries for running 32 bit applications. The amount of data to download for the drivers is really small compared to packages that contain CUDA libraries and tools. All packages have an higher Epoch set; so they should never be upgraded on your system when you enable this repository along the RPMFusion or ELRepo ones.

Package installation

If you are booting the system in UEFI mode; as a prerequisite to installing any external module (not built into the kernel package), you have to disable UEFI Secure Boot in the system configuration. All modules contained in the kernel package are signed with keys that are generated during build and deleted when packaging. If you want to preserve Secure Boot, you need to sign the modules yourself and import the keys into your hardware module. Doing so is out of scope here; if you need a decent guide just follow Red Hat’s guide for signing kernel modules.

First of all remove all the Nvidia drivers you might have on your system due to RPMFusion, ELRepo, or the Nvidia CUDA repository. The packages should already take care of this for you, as they should be completely compatible; but better be safe than sorry. This is usually accomplished with the following root command:

yum -y remove \*nvidia\*

Then, to install the Nvidia driver and its control panel utility in CentOS/RHEL with the binary kABI (Kernel ABI whitelist) module (this is the default on CentOS/RHEL), perform the following command:

yum -y install nvidia-driver nvidia-settings

To do the same in Fedora, perform the following command:

dnf -y install nvidia-driver nvidia-settings kernel-devel

Requirement on kernel-devel is required as otherwise the package kernel-debug-devel is pulled in automatically in place of the normal non-debug package. There is bug opened on dnf/libsolv for this.

Specific driver installations

For both Fedora and CentOS/RHEL distributions it’s possible to install additional packages and / or variant of the basic kernel modules. This paragraph contains some examples. Make sure you have the RPMFusion repository enabled if you plan to use akmod kernel modules on Fedora or EPEL if you plan to use DKMS modules on CentOS/RHEL.

akmod kernel module variant (Fedora):

dnf -y install nvidia-driver kernel-devel akmod-nvidia

DKMS kernel module variant (Fedora/CentOS/RHEL):

yum/dnf -y install nvidia-driver kernel-devel dkms-nvidia

To add 32 bit libraries on a 64 bit system (for games or applications like Steam):

yum -y install nvidia-driver-libs.i686

Additional driver configuration to your system

To add additional configuration to your system, just create the /etc/X11/xorg.conf file if it does not exist (by default it exists only in Red Hat Enterprise Linux / CentOS 6 systems). For example:

Section "Device"
    Identifier  "Device0"
    Driver      "nvidia"
    Option      "NoLogo" "true"
    Option      "DPI" "96 x 96"
    Option      "SLI" "Auto"
    Option      "nvidiaXineramaInfoOrder" "DFP-0"
    Option      "metamodes" "GPU-a493fbbb-7d76-86a2-8764-d76d487a75a7.DVI-I-1: nvidia-auto-select +0+0, GPU-c02960a4-be28-d5ce-8b02-be04b5e2550b.DVI-I-1: nvidia-auto-select +1680+0"
    Option      "BaseMosaic" "on"
EndSection

In this example we have 2 video cards with one monitor each, so we enabled SLI, Base Mosaic to have multi monitor support on SLI and make a layout with the second GPU monitor on the right of the first one. Also, we fix the DPI to 96×96, which is the hardcoded default in Gnome and in Open Source drivers.

Configuration for headless systems

Your system might only be used for CUDA development and not require the X server to be running the driver at all, so you might want to tweak the configuration a bit to make the system load (for example) the Intel driver as the main display and just the Nvidia driver for calculation. In this case, the Intel driver should load the modesetting driver, offload the rendering to the Nvidia driver and not use any monitor attached to the X server for the Nvidia driver.

So in this case, I would change /etc/default/grub to remove the nomodeset parameter to make the Intel KMS driver to load properly, regenerate the Grub config file, reboot and use this xorg.conf:

Section "ServerLayout"
    Identifier "layout"
    Screen 0 "intel"
    Inactive "nvidia"
EndSection
 
Section "Device"
    Identifier "nvidia"
    Driver "nvidia"
    BusID "<BusID for NVIDIA device here>"
EndSection
 
Section "Screen"
    Identifier "nvidia"
    Device "nvidia"
    Option "AllowEmptyInitialConfiguration"
EndSection
 
Section "Device"
    Identifier "intel"
    Driver "modesetting"
EndSection
 
Section "Screen"
    Identifier "intel"
    Device "intel"
EndSection

An example of the above can also be read in the official documentation.

The device file /dev/nvidia0 is normally created when loading the Nvidia driver, so if the X driver is not loaded the device file fileis not created. You can use the nvidia-modprobe command that is in the package with the same name.
It contains a SUID binary that creates the device files and set the appropriate permissions when automatic device creation is not available. It is called directly by Nvidia libraries:

$ for i in $(rpm -ql nvidia-driver-libs.x86_64); do
    strings $i | grep nvidia-modprobe > /dev/null && echo $i
done
/usr/lib64/libnvidia-cfg.so.1
/usr/lib64/libnvidia-cfg.so.381.22
/usr/lib64/libnvidia-eglcore.so.381.22
/usr/lib64/libnvidia-glcore.so.381.22
/usr/lib64/libnvidia-glsi.so.381.22
/usr/lib64/vdpau/libvdpau_nvidia.so.1
/usr/lib64/vdpau/libvdpau_nvidia.so.381.22

This requires some testing and adjustments with specifics to your setup, but is definitely possible to use the integrated Intel card and or rely on a system without X installed to run the CUDA components.

CUDA

Packaging

  • Previously in the repository was included the GPU Deployment kit. This was constructed with NVML (NVIDIA Management Library) headers, docs and samples from a separate tarball. The separate tarball was using a different version number than the drivers and was packaged in the nvidia-driver-NVML and nvidia-driver-NVML-devel packages. Installing these, the gpu-deployment-kit dependency provided by the CUDA repositories was preserved. Starting from CUDA version 8, the NVML header is provided by a CUDA subpackage (cuda-nvml-devel) and no longer provided as part of the GPU Deployment kit.
  • Included is also the NVENC (Nvidia Encoder) header, docs and code samples. Again, this uses a different version than the drivers.
  • All the libraries are split into subpackages, much like in the original Nvidia CUDA repository. This allows you to install and build software relying on specific components without the need to install all the CUDA toolkit just to satisfy a library dependency. With the new packaging organization, the original cuda-devel and cuda-extra-libs will pull in all the specific subpackages giving you the same situation you are accustomed to. Also, for the same reason, static libraries have been included in each respective devel subpackage.
  • In addition to the libraries bundled in the CUDA toolkit, also the cuDNN library for distributed neural networks is included in the repository. See the table below for details.

Distribution and CUDA version support

Operating systemel6 / el7f26 / f27 / f28
CUDA branch/version9.1.85.39.1.85.3
CUDA cuDNN version7.1.3.167.1.3.16
GCC compatibility-cuda-gcc (6.4.0)
Basic CUDA libraries/tools:

cuda
cuda-libs
cuda-extra-libs
cuda-cublas
cuda-cudart
cuda-cufft
cuda-cudnn
cuda-cupti
cuda-curand
cuda-cusolver
cuda-cusparse
cuda-npp
cuda-nvgraph
cuda-nvrtc
cuda-nvvp
YesYes
CUDA development:

cuda-cli-tools
cuda-devel
cuda-cublas-devel
cuda-cudart-devel
cuda-cudnn-devel
cuda-cufft-devel
cuda-cupti-devel
cuda-curand-devel
cuda-cusolver-devel
cuda-cusparse-devel
cuda-npp-devel
cuda-nvgraph-devel
cuda-nvml-devel
(also i686)
cuda-nvrtc-devel
YesYes
Java GUI programs:

cuda-nsight
cuda-nvvp
YesYes
Documentation and samples

cuda-samples
cuda-docs
(noarch)
YesYes

CUDA installations

To install just a runtime CUDA support (required for running CUDA enabled programs):

yum -y install cuda nvidia-driver-cuda

To install packages required for enabling CUDA development:

yum -y install cuda-devel

or, if you need to develop some application that requires multiple libraries:

yum -y install cuda-devel

A couple of examples. Just the basic tools:

$ sudo dnf install cuda
Last metadata expiration check: 0:00:20 ago on Sun Oct 23 13:11:01 2016.
Dependencies resolved.
================================================================================
 Package           Arch         Version               Repository           Size
================================================================================
Installing:
 cuda              x86_64       1:8.0.44-4.fc24       fedora-nvidia        95 M
 cuda-cufft        x86_64       1:8.0.44-4.fc24       fedora-nvidia        97 M
 cuda-curand       x86_64       1:8.0.44-4.fc24       fedora-nvidia        38 M
 cuda-libs         x86_64       1:8.0.44-4.fc24       fedora-nvidia       6.4 M
 
Transaction Summary
================================================================================
Install  4 Packages
 
Total size: 236 M
Installed size: 469 M
Is this ok [y/N]:

The basic tools along with all the libraries (note that the NVML headers are included):

$ sudo dnf install cuda-devel
Last metadata expiration check: 0:10:00 ago on Sun Oct 23 13:11:01 2016.
Dependencies resolved.
================================================================================
 Package                 Arch       Version             Repository         Size
================================================================================
Installing:
 cuda                    x86_64     1:8.0.44-4.fc24     fedora-nvidia      95 M
 cuda-cublas             x86_64     1:8.0.44-4.fc24     fedora-nvidia      21 M
 cuda-cublas-devel       x86_64     1:8.0.44-4.fc24     fedora-nvidia      38 M
 cuda-cudart             x86_64     1:8.0.44-4.fc24     fedora-nvidia     131 k
 cuda-cudart-devel       x86_64     1:8.0.44-4.fc24     fedora-nvidia     659 k
 cuda-cufft              x86_64     1:8.0.44-4.fc24     fedora-nvidia      97 M
 cuda-cufft-devel        x86_64     1:8.0.44-4.fc24     fedora-nvidia      73 M
 cuda-cupti              x86_64     1:8.0.44-4.fc24     fedora-nvidia     1.2 M
 cuda-cupti-devel        x86_64     1:8.0.44-4.fc24     fedora-nvidia     213 k
 cuda-curand             x86_64     1:8.0.44-4.fc24     fedora-nvidia      38 M
 cuda-curand-devel       x86_64     1:8.0.44-4.fc24     fedora-nvidia      60 M
 cuda-cusolver           x86_64     1:8.0.44-4.fc24     fedora-nvidia      23 M
 cuda-cusolver-devel     x86_64     1:8.0.44-4.fc24     fedora-nvidia     4.1 M
 cuda-cusparse           x86_64     1:8.0.44-4.fc24     fedora-nvidia      23 M
 cuda-cusparse-devel     x86_64     1:8.0.44-4.fc24     fedora-nvidia      23 M
 cuda-devel              x86_64     1:8.0.44-4.fc24     fedora-nvidia     1.6 M
 cuda-libs               x86_64     1:8.0.44-4.fc24     fedora-nvidia     6.4 M
 cuda-npp                x86_64     1:8.0.44-4.fc24     fedora-nvidia      91 M
 cuda-npp-devel          x86_64     1:8.0.44-4.fc24     fedora-nvidia      47 M
 cuda-nvgraph            x86_64     1:8.0.44-4.fc24     fedora-nvidia     4.6 M
 cuda-nvgraph-devel      x86_64     1:8.0.44-4.fc24     fedora-nvidia      12 k
 cuda-nvml-devel         x86_64     1:8.0.44-4.fc24     fedora-nvidia      41 k
 cuda-nvrtc              x86_64     1:8.0.44-4.fc24     fedora-nvidia     6.6 M
 cuda-nvrtc-devel        x86_64     1:8.0.44-4.fc24     fedora-nvidia      16 k
 
Transaction Summary
================================================================================
Install  24 Packages
 
Total size: 655 M
Installed size: 1.4 G
Is this ok [y/N]:

An example where your CUDA application just uses the CUDA Runtime API and not the kernel runtime:

$ sudo dnf install cuda-cudart
Last metadata expiration check: 0:13:10 ago on Sun Oct 23 13:11:01 2016.
Dependencies resolved.
================================================================================
 Package           Arch         Version               Repository           Size
================================================================================
Installing:
 cuda-cudart       x86_64       1:8.0.44-4.fc24       fedora-nvidia       131 k
 
Transaction Summary
================================================================================
Install  1 Package
 
Total size: 131 k
Installed size: 536 k
Is this ok [y/N]:

This will avoid you pulling in all the libraries as before just because you need a single library. This is useful for example for programs that leverage just some part of the CUDA toolkit, like the Nvidia Performance Primitives for image and signal processing in FFmpeg, and similar things.

Bugs

Just open an issue to the specific package on GitHub.

1,087 thoughts to “Nvidia driver, CUDA tools and libraries”

  1. Tried to update from version 396.24 to 396.45. On boot, X started in VGA mode with log showing
    NVRM: API mismatch: the client has the version 396.45, but
    NVRM: this kernel module has the version 396.24. Please
    NVRM: make sure that this kernel module and all NVIDIA driver
    NVRM: components have the same version.

  2. I am getting a strange issue where GDM does not start properly when a file /etc/X11/xorg.conf exists. I generated this in two ways: nvidia-xconfig (which places a default file there), and via nvidia-settings GUI which appears to be working properly otherwise. Perhaps there is another location where I need to write the config to? When GDM fails to start, you get a black screen without cursor. Switching to TTY2, or other using CTRL+F2, you see “A start job is running for Wait for Plymouth Boot Screen to Quit”. Booting as runlevel 3 (by modifying grub), removing the .xorg file, and rebooting results back to normal behavior.
    I’m trying to get a dual-monitor setup going, but it appears the system isn’t liking the xorg file? Any ideas?

      1. Do you mean nvidia-settings ? It detects the other monitor, but it doesn’t create a new X screen for it. So, nothing is displayed on the monitor?

        1. So, to change something, like adding a new X screen or changing the rotation of the monitor, it asks to save the configuration to the xorg.conf file. Hence, the issue. Any help appreciated.

    1. Unfortunately not, my crappy hosting provider does not allow me to 🙁
      Nothing prevents me to upload also somewhere else and use it as the master repo, though 🙂

  3. BTW, for anyone trying to get the CUDA rendering kernel built in Blender’s Cycles Renderer, and having its build attempts fail every time because it’s trying to use the system gcc (which on Fedora 28 is 8.1.1 currently, not supported even by CUDA 9.2):

    I discovered that blender does not respect the HOST_COMPILER environment variable, so even having that set by /etc/profile.d/cuda.sh won’t help with blender. And their official docs actually recommendediting the /usr/local/cuda/include/host_config.h header and removing the GCC version protection as a solution to that problem, which is just shockingly irresponsible.

    It’s also unnecessary. I discovered that blender does support a different, apparently undocumented variable (I had to read the blender source to discover it), CYCLES_CUDA_EXTRA_CFLAGS, which is used in the nvcc command line for the kernel build.

    So, if you install cuda-gcc and cuda-gcc-c++ from the negativo17 repos, and then launch blender as follows:
    CYCLES_CUDA_EXTRA_CFLAGS="-ccbin cuda-g++" blender
    …then the CUDA kernel will build successfully the first time Cycles tries to use your detected GPU as a rendering device, and you’ll be happily CUDA-rendering from that point forward.

    1. BTW, my system is a quad-core “Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz” with 8GB total system RAM, and I’m running a GeForce GT 710 with 1GB as my primary/only GPU and CUDA device, which barely meets the threshold to be considered a CUDA device at all. I think one of the CUDA 9.2 sample programs showed it as CUDA Capability 3.5.

      As a result of all that, while so far I’ve only rendered the default blender scene of a cube lit by a single light source, I’m pretty sure Cycles actually rendered it faster with my CPU than when using my GPU as the CUDA renderer.

      Not sure if that would hold up with more complex renders, I guess I’ll have to download some and time-trial it. But if you have a low-cost, underpowered GPU like mine, YMMV on the benefits of bothering with CUDA.

    2. Hi, thanks for the information, I did not know that the environment CYCLES_CUDA_EXTRA_CFLAGS could be passed at build time. There is also CUDA_NVCC_FLAGS which could be used, but I have some issues with CMake and variables with spaces in it (they get escaped by a backslash) so the list of parameters is not returned correctly.

      I just rebuilt Blender with CUDA support already compiled in, just install blender-cuda and it should be good to go: https://github.com/negativo17/blender/commit/dae27208b2b82eb8ee65cb03f0e5b2aadee487bf

        1. Oh, that’s unfortunate. Is there a way for 396.24.02 to maybe have a separate branch, like 396-vulkan? It includes certain necessary Vulkan extensions that are required by projects like DXVK, they simply don’t work with 396.24. Certain very popular Ubuntu PPAs even have this verison as default 396 package (https://launchpad.net/~graphics-drivers/+archive/ubuntu/ppa).
          I would happily manually install driver from Nvidia’s bash script, but for some reason Wine isn’t very happy about manually installed drivers, even regular 396.24.

  4. Thanks for a great package. I am not sure if my previous e-mail was successful.

    I am on an Optimus laptop with Nvidia and Fedora 28 (GCC 8). I am trying to build a
    very simple CUDA example. I have cuda-gcc and cuda-gcc-c++ installed.

    Can you please help me with this configuration?

    nvcc test1.cu
    /usr/include/c++/8/type_traits(1061): error: type name is not allowed

    /usr/include/c++/8/type_traits(1061): error: type name is not allowed

    /usr/include/c++/8/type_traits(1061): error: identifier “__is_assignable” is undefined

    3 errors detected in the compilation of “/tmp/tmpxft_000001b1_00000000-8_test1.cpp1.ii”.

    1. Just look at the output/man page of nvcc. You are just missing the -ccbin parameter.
      Look also at the cuda-samples Makefiles for examples.

      1. Thanks — can you update the Cuda notes above with cuda-gcc and ccbin? I’ll try this tonight, however I am now having trouble starting X (I’ll post more when I get back to the laptop at home).

        1. After installing cuda-gcc and cuda-gcc-c++ from negativo17’s fedora-nvidia repo, you can build all of the examples in CUDA 9.2 (installed from Nvidia’s own Fedora 27 repo) by going into the directory created by cuda-install-samples-9.2.sh and running HOST_COMPILER=cuda-g++ make.

    2. Have you managed to fix this?

      I have similar issue when compiling ethminer (Fedora 28, gcc 8.1.1, cuda 9.1):
      [ 37%] Building NVCC (Device) object libethash-cuda/CMakeFiles/ethash-cuda.dir/ethash-cuda_generated_ethash_cuda_miner_kernel.cu.o
      /usr/include/c++/8/type_traits(1061): error: type name is not allowed
      /usr/include/c++/8/type_traits(1061): error: type name is not allowed
      /usr/include/c++/8/type_traits(1061): error: identifier “__is_assignable” is undefined
      3 errors detected in the compilation of “/tmp/tmpxft_000008a3_00000000-14_ethash_cuda_miner_kernel.compute_70.cpp1.ii”.

      I think the reason is that version 8 of gcc is not supported by any version of cuda yet, see “Table 1. Native Linux Distribution Support in CUDA 9.2” at https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

      Searching Google for “/usr/include/c++/8/type_traits(1061): error: type name is not allowed” gives https://www.xpra.org/trac/ticket/1800?cversion=0&cnum_hist=4 where workaround was to switch to C++ 03 instead of C++ 11…

  5. Hi,

    I’m encountering the issue with black screen instead of GNOME with 396.24-1 driver.
    I had the same issue with the driver from rpmfusion and hoped that newer version of NVIDIA driver will solve the issue. But the issue is still there.
    This is what I see in the gdm-x-session in journalctl:
    Jun 07 19:55:09 zlopez-gamestation /usr/libexec/gdm-x-session[1896]: (EE) NVIDIA(GPU-0): Failed to acquire modesetting permission.
    Jun 07 19:55:09 zlopez-gamestation /usr/libexec/gdm-x-session[1896]: (EE) NVIDIA(0): Failing initialization of X screen 0

    There is a workaround however by booting to runlevel 3 and removing the kmod and let akmod build it again.

  6. Getting the following when running with no xorg.conf:

    [ 2983.335] (II) FBDEV(0): using default device
    [ 2983.335] (EE)
    [ 2983.335] (EE) Backtrace:
    [ 2983.336] (EE) 0: /usr/libexec/Xorg (OsLookupColor+0x13d) [0x59d5bd]
    [ 2983.336] (EE) 1: /lib64/libpthread.so.0 (funlockfile+0x50) [0x7f5644aac00f]
    [ 2983.336] (EE) 2: /lib64/libpciaccess.so.0 (pci_device_next+0x108) [0x7f564620bc18]
    [ 2983.337] (EE) 3: /lib64/libpciaccess.so.0 (pci_device_find_by_slot+0x3e) [0x7f564620bcbe]
    [ 2983.337] (EE) 4: /lib64/libpciaccess.so.0 (pci_device_vgaarb_init+0xaa) [0x7f564620d7da]
    [ 2983.337] (EE) 5: /usr/libexec/Xorg (xf86ConfigPciEntity+0x443d) [0x497bfd]
    [ 2983.337] (EE) 6: /usr/libexec/Xorg (xf86BusConfig+0x12d) [0x46c68d]
    [ 2983.338] (EE) 7: /usr/libexec/Xorg (InitOutput+0x96b) [0x47aebb]
    [ 2983.338] (EE) 8: /usr/libexec/Xorg (InitFonts+0x20d) [0x438d5d]
    [ 2983.338] (EE) 9: /lib64/libc.so.6 (__libc_start_main+0xeb) [0x7f56446fe18b]
    [ 2983.338] (EE) 10: /usr/libexec/Xorg (_start+0x2a) [0x42290a]
    [ 2983.338] (EE)
    [ 2983.338] (EE) Segmentation fault at address 0x0
    [ 2983.338] (EE)
    Fatal server error:
    [ 2983.338] (EE) Caught signal 11 (Segmentation fault). Server aborting

    If I put an xorg.conf, I get “no devices detected”.

    Any Ideas?

  7. I am getting a “(EE) modeset(0): drmSetMaster failed: permission denied” error on startup. If I switch to the nouveau, gdm starts up just fine. I am running Fedora 28 with a P2000 using kernel 4.16.13, 396.24 nvidia driver and dkms. Any ideas?

  8. Some update today stopped the driver from working. I just get a black screen instead of the display manager.

    I tried removing the nvidia-driver and nvidia-settings packages and the rest of the unneeded packages, rebooted and the system works fine with the nouveau driver

    Reinstalling the nvidia-driver, nvidia-settings and rebooting give me the black screen again.

    When I hit return, the black screen changes for a second to show the text mode bootup messages and that it is trying to start the display manager.

    Is anyone else seeing this on Fedora 28 x86_64 workstation with 5/23 updates?

      1. Are you using DKMS? There is an issue when DKMS is updated along with a kernel on the same update run. You can fix it by running these commands from a terminal, also when running under Nouveau:

        dkms remove nvidia/390.59 --all
        dkms install nvidia/390.59
        reboot

        Or use akmod-nvidia.

        1. No, I do not have dkms installed. I have akmod-nvidia-390.59-1 and kernel-4.16.10-300 on fedora 28. Before the updates yesterday that included that kernel, the nvidia driver was working fine.

          As a test, I tried booting the previous kernel and the nvidia driver doesn’t work for that kernel either. So maybe it’s the SELinux issue Dave linked below.

          There was another group of updates today to kernel 4.16.11-300. I applied these updates, rebooted and it didn’t make any difference.

          This is a very unusual situation as the nvidia driver you supply has been rock solid for a very long time. But the fedora folks have somehow managed to muck it up!

          1. Problem verified – downgrading selinux-policy-targeted fixes is a temporary fix.

            dnf downgrade selinux-policy-targeted

            An update to fix it is permanently being pushed.

          2. You mean an update to permanently fix that SELinux thing is under its way? How can we track that? And if I run that downgrade command, what happens? And then after the fix? Sorry, I’m a noobie about SELinux 😡

          3. I can verify – downgrading the selinux-policy-targeted worked for me. I switched from akmod to dkms, even removed and re-installed the drive from dkms. However nothing worked until I downgraded the selinux policy stuff. I didn’t have to reboot, just let X recycle. All is working for now. Thanks!

          4. @Gabriel Galli

            Yes, an update to fix the problem has been created and submitted to fedora updates. After appropriate testing, it will be pushed out in one of the future updates that fedora constantly gets when you run “dnf update”. In the meantime, you can continue using the nvidia driver by downgrading the selinux package. :”Downgrading” just means to install the previous version in place of the latest version.

            Whenever you run dnf update, it will update to the latest version and if that version is not one where the problem is fixed, you will need to downgrade the package again for the driver to work. It’s not a big deal – there is no downside.

          5. @Yaconsult ohh, okay, I got it now. I think I’ll wait for that update and then reinstall the driver. From the link that @Slaanesh posted, it seems they’ve already fixed it, it’s just not a public release yet. Thank you all very much! 🙂

    1. I’m seeing a similar issue. It seems for me as is the dkms-nvidia package is failing to install the kernel drivers properly. Fortunately my system automatically is falling back to nouveau driver.

      The dkms-nvidia package even seems to leave behind the /usr/src/nvidia-* directory it uses to build the kernel modules. If I build them by hand, I see a lot of missing signing/cert file warnings. I wonder if fedora changed how it does dkms secure package signing? This also seems to be a recent bug for virtualbox-dkms as well (on ubuntu though).

      1. There is an issue when DKMS is updated along with a kernel on the same update run. You can fix it by running these commands from a terminal, also when running under Nouveau:

        dkms remove nvidia/390.59 --all
        dkms install nvidia/390.59
        reboot

        Or use akmod-nvidia.

        1. Thanks that worked! One strange thing though… if I do a dnf reinstall dkms-nvidia it removes the dkms and then never reinstalls it… have to run the dkms install nvidia/390.59 command manually.

    2. Yeah, I’m experiencing exactly the same since yesterday. Boot halts at gdm – black screen.

      Not sure what the exact issue is, but if I move the two nvidia configs out of /etc/X11/xorg.cong.d/ then gdm starts OK, but the nvidia driver is not loaded – at least that’s what nvidia settings says. But I am able to run python scripts that make use of CUDA driver. Very odd.

      J.K.

      1. Are you using DKMS? There is an issue when DKMS is updated along with a kernel on the same update run. You can fix it by running these commands from a terminal, also when running under Nouveau:

        dkms remove nvidia/390.59 --all
        dkms install nvidia/390.59
        reboot

        Or use akmod-nvidia.

    3. Are you using DKMS? There is an issue when DKMS is updated along with a kernel on the same update run. You can fix it by running these commands from a terminal, also when running under Nouveau:

      dkms remove nvidia/390.59 --all
      dkms install nvidia/390.59
      reboot

      Or use akmod-nvidia.

    4. Same here but on Fedora Rawhide.

      The culprit being xorg 1.20.

      for now i downgraded xorg and upgrade my system using “dnf upgrade –exclude=xorg*” until the new nvidia driver is picked up by negativo, which supports xorg 1.20.

  9. Hello there, how you doing?
    So I’m running Fedora 28 and just received this 390.59 update, but now the GUI won’t load… 🙁
    Just tried reinstalling, but no luck. Boots normally if I leave it uninstalled and add “nouveau.modeset=0” to grub line.
    Oh, and also with the update came the kernel 4.16.10-300.
    Any ideas? Should I do anything after reinstalling, before trying to reboot?

    Best regards,
    Gabriel

    1. Are you using DKMS? There is an issue when DKMS is updated along with a kernel on the same update run. You can fix it by running these commands from a terminal, also when running under Nouveau:

      dkms remove nvidia/390.59 --all
      dkms install nvidia/390.59
      reboot

      Or use akmod-nvidia.

      1. No, I’m not using DKMS. Then I should install akmod-nvidia? I mean, now that I’ve uninstalled your driver, I should reinstall it and install akmod, is that correct?

  10. Is there somewhere I can find an explanation of what the “long lived” and “short lived” drivers are?

    I’ve tried googling but the information is a bit contradictory

    The table here says fc28 has both the “short lived” and the “long lived” ones. I take this to mean that there should be two different versions available, but when I use it on my fc28 install I only see 390.48, is that “long lived” or “short lived”? Is it both?

    The nvidia website lists 390.59 and 396.24. Are these “short lived” or “beta”?

    I’m very confused by this versioning scheme

  11. After installation of nvidia-driver-390.59-1.fc28 on x86_64 Linux 4.16.9-300.fc28.x86_64 dmesg tells:
    quote
    Bad or missing usercopy whitelist? Kernel memory exposure attempt detected from SLUB object ‘nvidia_stack_cache’ (offset 11440, size 3)!
    [ +0.000008] WARNING: CPU: 5 PID: 1127 at mm/usercopy.c:81 usercopy_warn+0x7d/0xa0
    unquote
    What does this mean?
    What to do?
    Rgds
    AW

  12. Hi, quick question. I have installed already nvidia-driver, cuda-devel, cuda-cudnn followed by http://blog.mdda.net/oss/2016/11/25/nvidia-on-fedora-25.
    The problem is i’m trying to compile “darknet” (DL library) with cuda. it needs to specify the folder /usr/local/cuda that is made when installing with cuda on nvidia website. But i can’t find them. Thanks in advanced.

      1. Many thanks, I have compiled darknet successfully. I’m trying to map the negativo cuda lib to a cuda folder to easier to use for further usages.

        1. As a general rule, you can point the original “CUDA_HOME” variable to “/usr” and just adjust where the program is looking for the headers (/usr/include/cuda). All the rest should be fine.

  13. BTW is there any other channel for bug reporting or discussions in general? Replying to this thread is not the best choice IMHO. Even a google-group mail list would be an improvement, I believe. Just my $0.02 😉

  14. Hi,

    it seems there is some sort of conflict between Negativo’s and RPMFusion’s packages:

    Last metadata expiration check: 0:03:15 ago on ter 22 mai 2018 09:40:09 -03.
    Dependencies resolved.

    Problem: package nvidia-driver-3:390.48-1.fc28.x86_64 conflicts with xorg-x11-drv-nvidia provided by xorg-x11-drv-nvidia-3:390.59-1.fc28.x86_64
    – package akmod-nvidia-3:390.59-1.fc28.x86_64 requires nvidia-kmod-common >= 3:390.59, but none of the providers can be installed
    – cannot install the best update candidate for package nvidia-driver-3:390.48-1.fc28.x86_64

    – cannot install the best update candidate for package akmod-nvidia-3:390.48-1.fc28.x86_64

    Package Arch Version Repository Size

    Skipping packages with conflicts:
    (add ‘–best –allowerasing’ to command line to force their upgrade):
    xorg-x11-drv-nvidia x86_64 3:390.59-1.fc28 rpmfusion-nonfree-updates 2.5 M
    Skipping packages with broken dependencies:
    akmod-nvidia x86_64 3:390.59-1.fc28 rpmfusion-nonfree-updates 74 k

    Transaction Summary

    Skip 2 Packages

    Nothing to do.
    Complete!

    @Slaanesh: just to make sure you are aware. Should we just wait a little longer, or is there any other recommended course of action?

    Thank you as always for all the hard work!

    1. The packages have been updated. If you plan to use RPMFusion packages as well, I suggest you just switch to their Nvidia driver.

      1. Hi Slaanesh, thank you for your reply, and for the fix. I understand there might be some conflicts, but, is it really necessary to disable all RPMFusion repos? I am interested in some other packages they provide. Is there any less drastic solution? Like, for example, raising the priority for your packages over RPMFusion’s?

        I have all these config files with enabled=1 on /etc/yum.repos.d:

        rpmfusion-free.repo
        rpmfusion-free-updates.repo
        rpmfusion-nonfree.repo
        rpmfusion-nonfree-updates.repo

        And I also have rpmfusion-nonfree-nvidia-driver.repo, but it is disabled.

        1. Probably not, as there might be conflicts and dependency issues. If you need other packages in RPMFusion, I suggest you to use RPMFusion directly and disable the repositories hosted here. My resources are limited and at the moment I’m interested in maintaining a definite set of packages.

  15. Since kernel 4.16.6+ (I am currently running 4.16.9) on Fedora 28 my Gnome session no longer opens with nvidia drivers installed from negativo repository. Prior, with Fedora 27 there was no issues at all. On Fedora 28 with kernel 4.16.3, I had to uninstall/reinstall the nvidia-driver along with the dkms library after each halt in order to have a running session. Now, even that is no longer working.

    In journalctl, I noticed the following message:

    nvidia: module verification failed: signature and/or required key missing – tainting kernel

    If I understand, UEFI should be disabled but I never needed that previously. Is it a new requirement?

    Please note that removing the nvidia driver (thus using the nouveau driver) resolves the problem but its not a viable solution since I need to use some applications that requires the nvidia driver.

    Any other person with the same issue?

    1. If I understand, UEFI should be disabled but I never needed that previously. Is it a new requirement?

      UEFI should not be disabled, I guess you are confusing with Secure Boot on UEFI. Secure boot must be disabled, as the modules are not signed and the original signing keys of the kernel (and built in modules) is a throw away key that is not usable. Unless you plan to manually sign the modules yourself, I would suggest to disable Secure boot.

  16. nvidia-libXNVCtrl.i686 seems to have been dropped from F28 repo, was this for a reason or by mistake?

    1. Actually it was planned. For Fedora 28 I just pushed x86_64 builds with i686 multilibs only; and nobody complained so far. Also the latest driver is 64 bit only.
      nvidia-libXNVCtrl.i686 is used only by nvidia-settings.i686, why would you need it on a x86_64 system? I can re-add it, but I don’t see the point.

      1. No, not needed. I had it installed just in case, but I confirmed your findings that its only used by nvidia-settings. Thanks for the confirmation

  17. Linux driver 396.24 has dropped NVS 5400M support. Do you have any plans to issue a “legacy” driver like 396.18 (beta) which supported NVS 5400 chips.

  18. I have laptop with intel and nvidia graphic card when i install akmod-nvidia, nvidia-drivers etc. system starts with nvidia graphic as default. I use fedora 28 in fedora 27 system was started with intel as default and nvidia as discrete.
    What im doing wron or where i can set default card ?

    1. Hello, not yet (it has been released today!). Being a new official release, not long lived, I will update all Fedora versions with it as per the table in the page.

    1. All packages for Fedora 28 along with the latest updates should be there, with the exception of gstreamer1-plugins-bad, which will be probably pushed tomorrow.
      There is not yet a working cuda-gcc package as I was not able yet to create a working GCC 6 on Fedora 28 (so ccminer and blender are also missing.

      1. This basically means any CUDA development on F28 is not possible now – I have just ran into this issue :/… CUDA version used here is 9.1 which does not work with gcc 8…

Leave a Reply to slaanesh Cancel reply