How to uninstall CUDA9.0 and cuDNN under Win10?

I have installed CUDA9.0 and cuDNN9.0 under win10, but tensorflow1.4 does not support them in this version.
How to uninstall them completely?

relevant programs in my control panel
NVIDIA Tools Extension SDK
NVIDIA Nsight Visual Studio Edition
NVIDIA CUDA Visual Studio Integration
NVIDIA CUDA Samples
NVIDIA HD Audio
NVIDIA PhysX
NVIDIA GeForce Experience
NVIDIA Device Driver
NVIDIA 3D Vision
NVIDIA CUDA Runtime
NVIDIA CUDA Documentation
NVIDIA CUDA Development

Thanks a lot

4 Likes

regarding CUDA 9 uninstall:

[url]Installation Guide Windows :: CUDA Toolkit Documentation

Through the Program and Features widget in control pannel, I uninstalled:

NVIDIA Nsight Visual Studio Edition
NVIDIA CUDA Visual Studio Integration
NVIDIA CUDA Samples
NVIDIA CUDA Runtime
NVIDIA CUDA Documentation
NVIDIA CUDA Development

But, again if I try to install NIVIDIA toolkit by running (cuda_9.1.85.3_windows.exe), it tells me that the tool is installed. Hence, how to fix this issue?

Thanks in advance

1 Like

Hey all. I ran into the same problem but 3 months later.

I SOLVED IT!!!

I needed CUDA 8 installed for CAFFE to build. I also made the mistake of installing the latest NVIDIA (9 and 9.2).

The fix is not 100% known, it might be all of these steps or just the last couple of steps:

  1. I uninstalled 9.2
  2. I uninstalled 9 - NOTE - The Visualiser was still there, and all files were in the NVIDA folders (under Program Files).
  3. I installed 8.2 (service pack/update).

At that point I was able to compile Caffe.

Now I’m going to install 9 and 9.2 on again. Best of luck

@pleitch When you say you uninstalled 9.2 and 9 how did you go about that?. Did you uninstall from the control panel or through some other means?

1 Like

I was hitting this issue; it looks like at some point some installer metadata was removed.

I went looking for the uninstaller info in the registry.

For example:

HKEY_LOCAL_MACHINE\SOFTWARE\Microsoft\Windows\CurrentVersion\Uninstall\{B2FE1952-0186-46C3-BAEC-A80AA35AC5B8}_cublas_10.0

Has a reference to the file:

C:\Program Files\NVIDIA Corporation\Installer2\installer.{DA2993F0-A753-4637-91EE-8738FAA315EB}\setup.CFG

However, on my system this file didn’t exist (anymore).

The solution was to re-install the exact version I was trying to remove and then uninstall the components. During the installation use “custom” rather than “express” and pay special attention to the version numbers to avoid rolling back new driver components.

My best guess is that some tool blasted away the C:\Program Files\NVIDIA Corporation directory any any installer metadata that was under it.

Hi …
It is easy …
Just opens control panel/programs and uninstall theses:

  1. NVIDIA CUDA Runtime "version"
  2. NVIDIA CUDA Samples "vertion"
  3. NVIDIA CUDA Documentation"vertion"
  4. NVIDIA CUDA Developement "vertion"
  5. NVIDIA CUDA Visual Studio Integeration "vertion"
3 Likes

I tried this but the it’s not uninstalling cuda 11.
I want to keep only cuda 10.1

2 Likes

I had not installed VS2019 prior to the first install, so I wanted to uninstall and reinstall the CUDA toolkit, but the Windows “Add or Remove Program” didn’t work so effectively. Eventually I redownloaded the Toolkit exe file and just ran it. It uninstalled any old versions and installed the new version. Hope it works for you too…

1 Like

I have the same problem. I uninstalled everything related to NVIDIA in control panel, cuda 11.1 was installed (I’m on a windows 10 machine). I then downloaded CUDA 10.2 toolkit and installed it. When I open control panel now it only shows 10.2. But when I open cmd and type: nvidia-smi, it says I have version 11.1 installed.

5 Likes

nvidia-smi is part of the driver install, not the toolkit, so if you are using a later than 10.2 driver version, it, (and the driver), will have been compiled with a later toolkit. I believe this is what the version number in the nvidia-smi output shows, not the installed toolkit.

1 Like

same problem is showing on my system. didi you resolve it?

2 Likes

no its false!

I have the same problem but on windows 11 and with Cuda toolkit 12.x
How did you resolve it?