CUDA Driver Issue?

My installation is not working with my current Drivers, I am using version 390.65 and CUDA 8.0 GPU is a 1070 on windows. How do I fix this? When I try to install I get the error “This NVIDIA graphics driver is not compatible with this version of windows.” My windows is currently up to date. I can not switch to CUDA 9.0 because I am using tensorflow. Any help would be appreciated!

It’s been a few months since I last installed Tensorflow on a Windows box, but I recall the (Tensorflow) doc said to install a driver that’s appropriate for CUDA 8.0, which was not the latest-and-greatest driver (I felt somewhat offended by that). When installing the full CUDA dev kit package, it includes drivers. So I think the first question to ask is whether that 390.65 driver you’re running was installed by the CUDA 8.0 installer, or if it was installed after-the-fact. If it is a later driver than CUDA 8.0 bundles, rolling back to the drivers that are packaged with the current 8.0 release is probably a good thing to try.

Best of luck!

= B

I rolled back my drivers, and still no luck. Same error as before

This is not happy time for people installing CUDA drivers.
The night I answered your post I was up literally until 8AM (having started around 830PM the previous night) simply trying to do a clean install of CUDA on a new computer with 1080Tis and UBUNTU 16.04 - which I have done before without much trouble.
I ran into four separate sets of problems, each of which masked the next one - total freakin nightmare.
In fairness to Nvidia, half the blame goes to Canonical / Ubuntu, who need to face the reality that their open-source Nouveau Nvidia drivers don’t even pretend to support CUDA, and they need to support the user community with deb packages that have Nvidia proprietary binary drivers baked in.

In your case there may be an outside-the-box solution: if you’re willing to learn R, which IMO is far better suited to ML than python, RStudio maintains R Tensorflow packages that are much better packaged and easy-to-use than Google packages for other “officially supported” languages. The R package installer system itself makes sure all the GPU stuff is properly installed and configured. Full Disclosure: I have not tried this myself because I’m not so hot on TF anymore, however my friends who have done so say it’s amazing. Of course that could all be true and the RStudio version could be hosed just like everyone else is right now.

  • B