Pytorch with cuda9.2 10.1 10.2, what's the best in most situations?

If we have the latest Nvidia driver, we can install pytorch with cuda in 3 ways:
conda install pytorch cudatoolkit=9.2 -c pytorch
conda install pytorch cudatoolkit=10.1 -c pytorch
conda install pytorch cudatoolkit=10.2 -c pytorch

All three methods work. So what is the difference among installed pytorchs with 3 types of cuda? Such as running efficiency or features?

Which one should we choose in most situations?

use the latest, i.e. 10.2

1 Like

Could you briefly describe the benefits of using cuda10.2?

in a nutshell, you get more optimizations, fixes, etc.
sometimes newer versions interduce some bugs, in which case, you check and if that affects you, you use the former (bug-free) version.
in any case, as a rule of thumb try to use the latest version unless you have a good reason not to (compatibility issues, etc in your environment or pipeline for example)

Also you can always check the changelog and see whats new/different compared to the former version.