I am trying over the last 6 hours to install CUDA on Ubuntu 17.04 Gnome with an NVIDIA GeForce 8800 GTS 320 MB but with no success. I tried the following:
- I installed the recommended driver for the aforementioned card but all I get is artifacts and the distro stops working at a certain point.
- I installed later drivers with no luck as I get a flashing screen and low resolution on the login logo.
Do you have any idea what could I do so I can have the correct driver, CUDA version and a working OS? I pretty much tried many things that I found here and around the web but again no luck at all.
If there is something I missed here on askubuntu which solves my problem please point it out and mark as duplicate my question. ( I am sorry for the duplicate if it exists :) .)
I suggest you set up the recommended Nvidia drivers and get them working before installing the cuda-repo-ubuntu1604-8-0-local-ga2_8.0.61-1_amd64.deb package from Intel.
Get the drivers from the Software Updater/Settings button/Alternative drivers tab, not from Intel directly or from the CUDA package. Use the recommended (tested) version of the latest driver. You could try an earlier version to maybe avoid artifacts, but I never bothered. If you got the drivers elsewhere, uninstall them and redo the installation as described above.
Installing the Intel CUDA package will bring in all the other cuda packages in the (multiverse) repositories (if you've enabled the multiverse in the software update settings). This CUDA release requires the gcc-5 and g++-5 compilers, so since gcc-6... is the default for 17.04, install them manually. I just relink the /usr/bin/gcc and /usr/bin/g++ to the "5" versions instead of the "6"s, but I guess you could fiddle with the makefiles if you want to keep the "6" versions as the default. See if the /usr/local/cuda/bin got added to your PATH at the front, (log out and in again). If not either set it up when you login, or just have a script to set it when you work with CUDA. Same for adding /usr/local/cuda/lib64 to LD_LIBRARY_PATH. Be sure to export these variables after they are set.
Go to the writeable cudasamps directory (it may show up in your home directory). If it's not in your home dir, then copy it yourself from the /usr/local/cuda installation to some writeable (by you) location. Try to run a makefile from one of the sample directories -- e.g. ...cudasamps/5_Simulations/nbody The makefile should complete and produce the nbody executable, try to run it. That's about it. The Nvidia drivers I've used for this were the 367 and 375 versions (yes with the windows artifacts after sleep for the 375 driver).