How to create a clean install of CUDA 10.0

Platform: Win 10 64-bit Home

Months ago I installed CUDA 10.1 (with 10.0, 9.0 and ?9.2 installed) in anticipation of TensorFlow2 but it turns out that TF2 needs 10.0 and I can’t get back…

The system is in a confused state after uninstalling 10.1, 9.x (2 off); despite the uninstalls and re-installaton of CUDA 10.0 (at least twice - with reboots) - nvidia-smi reported - that the nvcuda.dll was still 10.1; tracked it down to system32 and replaced it (after rename) with nvcuda64.dll from driverstore\filerepository and now nvidia-smi reports no CUDA driver at all.

This question https://stackoverflow.com/questions/58235018/cuda-issue-how-to-clean-install-cuda-in-win-10-to-resolve-cudagetdevice-fail on StackOverflow has gory detail for the morbidly curious…

All I would like to know is how to start from scratch and do a clean CUDA 10.0 installation?

Many thanks!

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 436.48       Driver Version: 436.48       CUDA Version: N/A      |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 108... WDDM  | 00000000:01:00.0  On |                  N/A |
|  0%   37C    P8    23W / 225W |    954MiB / 11264MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

nvidia-smi said no CUDA but I thought I’d try since the nvidia control panel DID say there was 10.0 … I got “The procedure entry point fatBinaryCtl could not be located in the dynamic link library C:\WINDOWS\SYSTEM32\nvcuda.dll”

It’s normally not necessary to clean everything off your system. However a fairly full set of instructions for complete removal of NVIDIA software is here:

https://devtalk.nvidia.com/default/topic/1033111/cuda-setup-and-installation/cuda-9-1-cannot-install-due-to-failed-visual-studio-integration/post/5259359/#5259359

step D would be slightly different of course if you are installing CUDA 10.0

@Robert_Crovella - thanks; I shall dedicate tomorrow to trying all that (I agree with others in that thread that there should be a better way from Nvidia) but if that’s what it takes…

But, when you say “It’s normally not necessary to clean everything off your system.”, given the issues I noted (and the extra info on S.O) do you have something else you would recommend? I thought I had tried everything else…

And I shall continue to hold out hope for less arduous suggestions from others :)

Cheers, J

PS I just noted on this TF page https://www.tensorflow.org/install/source#tested_build_configurations that the tested builds use cuDNN 7.4 and I have been using either what the 10.0 installer put there or 7.6.3? Should I first try 7.4? If so, which one?

Version Python version Compiler Build tools cuDNN CUDA
tensorflow_gpu-1.14.0 2.7, 3.3-3.7 GCC 4.8 Bazel 0.24.1 7.4 10.0

A simple removal of cuda toolkits using instructions provided in the windows install guide should be sufficient for removal.

Thereafter, install the CUDA toolkit of your choice. It’s probably best to install and keep the latest driver for your GPU (e.g. your 436.48 driver – that will work with any currently available CUDA version). When installing the CUDA toolkit of your choice, deselect the option to install the bundled driver.

BTW, if you have manually moved around and renamed dll’s in the windows system folders, I would say all bets are off. Who knows what is up with that machine. I would not be able to sort that out. My best advice would be to reload windows. I wouldn’t be able to expect any tools from NVIDIA to be able to unwind or sort out such things.