TL;DR: What is the best way, using built-in Windows Server tools as a local administrator, to restrict access to GPU 0 on a Windows Server workstation with two Nvidia GPUs to a select group of users by setting CUDA_VISIBLE_DEVICES?
I have the following configuration: Two Nvidia GPUs are installed on a Windows Server workstation. Multiple users connect via Remote Desktop to perform their computations on the workstation using their workplace domain accounts. One GPU, GPU 0, has better capabilities and is supposed to be reserved for certain users who actually make full use of it. Other users with less GPU-intensive tasks shall be prevented from (unknowingly) blocking GPU 0. Instead, they should only be able to use the second, less powerful GPU, GPU 1.
I would be fine with a solution where users only see GPU 1 by default and I have to explicitly grant access to GPU 0.
Note that I only have local administrator privileges and need to work within that constraint.
I have tried setting environment variables for users in the registry as described here. However, this is very cumbersome as you can only make changes for logged in users, which is not recommended as described in this post . Also, you would have to repeat this for each new user.