I used old Fedora 25 to use cuda toolkit 9.1.85. However when I installed nvidia driver 384.98 and cuda toolkit 9.1.85 on Fedora 25, boinc-client service failed to recognize my GPU. It says NVIDIA driver is present, but no GPUs found…
No. I haven’t programmed yet. I only checked nvidia driver. glxinfo says direct rendering is possible, so the driver looks good. However since boinc core says it doesn’t detect GPU, nvidia driver can’t accept cuda/opencl application. Are there any good confirmations?
Rebooting didn’t help.
nvidia-smi told me the status of nvidia gpu (how much load imposed on and temperature and so on). So it look like working. My driver may be working. boinc is a bit wrong. I will try Ubuntu.
I just installed Nvidia Cuda 9.1 and tried to run Tensorflow on an Ubuntu 17.10 distro with an nvidia-384 driver. When I tried to run a TF example I got the message that the Cuda version is later than the driver can handle.
So I believe that a later driver will be needed, which is too bad, because the Nvidia driver situation on Ubuntu linux is abysmal, and every new driver usually requires 1 or more days futzing around just to get back to a login screen. I just spent a day trying to get Ubuntu 17.10 to work with my new (1080ti) computer and finally had to switch to UbuntuMate, which seems to be the only distro that works right from the start with Nvidia cards. Now I have to wait for the Mate folks to get a 387 or 388 driver into their build.
You may want to backpedal to Cuda 9, which I think is the latest that 384 drivers can recognize.