I recently got the new Geforce GTX 1080. It is a monster. Boasts 8 gigs of memory and 2500 cuda cores. I am using it for my deep learning experiments.
I had an existing GTX 980 Ti which I use for display. I already had nvidia driver 343 installed from
apt-get and cuda-7.5 running fine. driver-343 is actually old and not compatible with GTX 1080.
It is relatively pain free to physically install GTX 1080 on your PC. Be sure to have a power supply at-least 600W.
Basically these are the steps to follow:
- Install nvidia driver 367
- Install cuda (An important pitfall with this, see below)
Install driver 367
Warning : Do not trust apt-get on this. Their drivers are almost always out-dated
Get the latest driver from [Nvidia-drivers]. I prefer the .run file nvidia provides. I know a couple of years ago it used to cause issues with your X-system. Seems like it is now polished and works well.
Uninstall previous nvidia drivers.
$ sudo apt-get purge nvidia-*
Stop light gdm (graphical interface)
$ sudo service lightgdm stop
Go to tty (CTRL+ALT+F1). Set your init state to 3 (text only mode). It is important to do this. Note these commands on a paper or something. I experienced sometimes the tty does not show with the newest driver. I just ssh to my PC as a way around.
$ sudo init 3
Log in to tty and cd to the directory where your have downloaded the driver.
$ sudo ./NVIDIA-Linux-x86_64-367.35.run
It will ask if you want to install 32-bit libraries, say no (assuming you do not have a 32-bit OS, hopefully. If you do have a 32-bit OS it is a good idea to upgrade…)
In a few minutes it is done….smooth. Reboot your PC
$ sudo reboot
As a test if the driver worked, try running nvidia-smi and make sure
Although GTX 1080 can work with cuda-7.5, it is worth while to go for cuda-8 as it will be able to take advantage of the tesla micro-architecture of the GTX 1080. Also apparently there were improvements in the cuda compiler (nvcc). If you be lazy and install cuda with apt-get you get cuda 7.5. In short go for cuda-8.
Go to cuda official page [Nvidia-CUDA] and download the .run file for cuda-8. Initially I had attempted this with the .deb and had issues with display. I recommend using .run file.
Before proceeding ahead, ensure you have driver 367 correctly installed.
Navigate to the directory where you downloaded the .run for cuda 8. Execute the .run
$ sudo cuda_8.0.27_linux.run
It asks you to install driver 361 (say NO to this. This is very important. Driver 361 is not compatible with GTX1080. Driver 367 is compatible).
After you say no to this you answer a few more questions (say yes to all) and it installs cuda in /usr/local/cuda-8.0.
$ export PATH=”$PATH:/usr/local/cuda-8.0/bin/:”
That is it. You may like to see if the CUDA-samples work. Once you set the PATH you could run the following to copy samples to your home directory
$ cuda-install-samples-8.0.sh ~/$ cd NVIDIA_CUDA-8.0_Samples/
Feel free to run more examples and be satisfied that your new graphics card is working with cuda-8.0 and the newest driver.
Good luck and comment to discuss/clarify more.
Recently got my hands on Nvidia Titan X (Oct 2016). Failed to installed drivers with this process. Turns out it was due to come kernel config issues that I cannot install from Nvidia’s binaries. Do not understand this problem in detail.
However, I would used the Ubuntu ppa and installed the latest driver (nvidia-driver 370).
It is extremely important to either Turn off the UEFI Secure boot from your BIOS. NVidia driver doesn’t load with UEFI secure mode on. Secure boot is not available in a lot of motherboards. But the high end ones will have it. If you have an Titan X, you also possibly have a high end mother board, so beware.