So, maybe you have decided to get involved in this new Deep Learning wave of open source projects. The neural networks are kind of slow on a traditional computer. They have to do a lot of matrix math across thousands of neurons.
The traditional CPU is really a latency engine...run everything ASAP. The GPU, on the other hand, is a bandwidth engine. It may be slow getting started but it can far exceed the CPU in parallelism once its running. The typical consumer CPU is 4 cores + hyperthreading which gets you about 8 threads (virtual cores). Meanwhile, an entry level Pascal based GeForce 1050 will give you 768 CUDA cores. Very affordable and only 75 watts of power. You can go bigger but the smallest is huge compared to a CPU.
I've looked around the internet and haven't found good and complete instructions on how to setup for an Nvidia video card on a current version of Fedora. (The instructions at rpmfusion is misleading and old.) So, this post is dedicated to setting up a Fedora 25 system with a recent Nvidia card.
With your old card installed and booted up...
1) Blacklist nouveau
# vi /etc/modprobe.d/disable-nouveau.conf
add the next line:
2) Edit boot options
# vi /etc/default/grub
On the GRUB_CMDLINE_LINUX line
save, exit, and then run either
# grub2-mkconfig -o /boot/grub2/grub.cfg
Or if a UEFI system:
# grub2-mkconfig -o /boot/efi/EFI/<os>/grub.cfg
(Note: <os> should be replaced with redhat, centos, fedora as appropriate.)
3) Setup rpmfusion-nonfree:
# wget https://download1.rpmfusion.org/nonfree/fedora/rpmfusion-nonfree-release-25.noarch.rpm
# rpm -ivh rpmfusion-nonfree-release-25.noarch.rpm
4) Enable rpmfusion-nonfree
# vi /etc/yum.repos.d/rpmfusion-nonfree.repo
# vi /etc/yum.repos.d/rpmfusion-nonfree-updates.repo
In each, change to:
4) Update repos
# dnf --refresh check-update
See if new release package
# dnf update rpmfusion-nonfree-release.noarch
5) Start by getting rid of nouveau
# dnf remove xorg-x11-drv-nouveau
6) Install current nvidia drivers:
# dnf install xorg-x11-drv-nvidia-kmodsrc xorg-x11-drv-nvidia xorg-x11-drv-nvidia-libs xorg-x11-drv-nvidia-cuda akmod-nvidia kernel-devel akmod-nvidia --enablerepo=rpmfusion-nonfree-updates-testing
7) Install video accelerators:
# dnf install vdpauinfo libva-vdpau-driver libva-utils
8) Do any other system updates:
# dnf update
9) Shutdown and change out the video card. (Note that shutdown might take a few minutes as akmods is building you a new kernel module for your current kernel.) Reboot and cross your fingers.
This should get you up and running with video acceleration. This is not a CUDA environment for software development. That will require additional steps which involves registering and getting the Nvidia CUDA SDK. I'll leave that for another post when I get closer to doing AI experiments with the audit trail.