Installing CUDA consistently breaks graphics drivers (have tried multiple clean installations)

Question

I have been trying to install CUDA Toolkit, but I have been running into the same issues even after multiple clean Ubuntu installations. After reboot, I run into a login screen with incorrect resolution. Logging in quickly breaks back to the same login screen. Here is the issue.

One possible duplicate issue. And another. Neither has a satisfactory answer for how to properly install CUDA.

Here is my process:

1. Fresh installation of Ubuntu 16.04 from USB drive on clean partition
2. Download a couple applications (Intellij, Conda, etc)
3. Verify GPU with lspci | grep -i nvidia
4. Check distribution with uname -m && cat /etc/*release
5. Check gcc installation with gcc --version
6. Check kernel with uname -r
7. Double-check kernel with sudo apt-get install linux-headers-$(uname -r) (no updates found)
8. Download CUDA for Linux x86_64 Ubuntu 16.04 deb (local)
9. Check checksum with md5sum <file>
10. No conflicting installation methods should exist (???)
11. sudo dpkg -i cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb
12. sudo apt-key add /var/cuda-repo-<version>/7fa2af80.pub
13. sudo apt-get update
14. sudo apt-get install cuda

The installation finishes properly and I reboot the computer to the above issue. What am I doing wrong? What is a workaround? I'm fine with reinstalling Ubuntu one more time as long as I can get CUDA properly installed.

Answer 1

On afresh install, get the Ubuntu provided Nvidia drivers installed and working before doing the CUDA install. After the install of the cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb, install the Ubuntu cuda package, which gets the toolkit.

---

From the Dash/System Settings, under the System section, click on the "Software & Updates" icon, and check that under the "Ubuntu Software" tab, the "Proprietary drivers for devices (restricted)" is selected, then under the "Additional Drivers" tab, select the Nvidia binary driver that's tested (first one?). Click on the "Apply Changes" button on that window, and then restart the X server, or just log out and in again. Check the running driver with the "nvidia X server settings" icon from the dash. Reboot just to check again, and if still running the Nvidia server, and no login problems (see many solutions here for login loops -- basically clean up old .Xauthority or .cache or .config files in your home directory) you are all set to install CUDA.