ubuntu 17.04 + nvidia 281 + cuda 8.0 = cuda driver version insufficient?

Hi, I have a GTX 1080 ti.

Installed the CUDA-8.0 from https://developer.nvidia.com/cuda-downloads
Used local deb option.

Upgraded the cuBLAS with the patch available on the same page.

Re-installed the nvidia driver version 281 using the ppa:graphics-drivers/ppa (the CUDA 8.0 install removed it and put 375 instead, which does not work with 1080 ti).

I’ve managed to compile most of the samples using gcc 6.3 with export EXTRA_NVCCFLAGS="-Xcompiler -std=c++98"

(3_Imaging/cudaDecodeGL failed due to not finding -lnvcuvid - but deviceQuery was built successfully)

When I run deviceQuery, though, I get the following:

./deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

Any idea how to go about and fix this?

Use a supported configuration for CUDA 8. (read the NVIDIA CUDA install guide)
Follow the instructions in the CUDA install guide, carefully. I suggest the runfile installer method, since it gives easier control over driver selection.
The only exception being, install the proper driver for the 1080Ti first (it is not driver version 281, and get your driver from NVIDIA, not ppa), and when using the runfile installer method, deselect the option to install the driver that comes with the CUDA toolkit installer.

Oops I meant to say driver version 381, not 281… Isn’t it the latest driver? Do I need a different one?

You have to use 367 for a gtx 1080ti, cuda 8 and gcc 5.3.1 (ubuntu 16.04).

The new drivers work if you use cuda 9 (deb, not run file) on ubuntu 17.04 (which uses gcc 6.3.0)

367 is not the right driver branch for CUDA 8 anymore (i.e. CUDA 8.0.61), nor is it the correct branch for GTX 1080Ti