New CUDA 11.x versions add support for TF32 format, other new features for newer cards (RTX30xx, A100 etc).
Is it worth upgrading to CUDA 11.x if you have GTX 1050 or RTX 2080 (having tensor cores)?
Could it be that new features only add computational overhead (at least in the size of installation file, they do), and an older GPU won't be able to use the new features?
Topic cuda deep-learning
Category Data Science
There is no right answer to this question, because there are many factors to consider, for instance:
Indirect dependency on specific/minimum CUDA version: with deep learning frameworks like pytorch and tensorflow, a specific version of the framework depends on a specific version of CUDA/cuDNN, so the decision of which (minimum) CUDA version to have is directly dependent on the framework version you need.
Bug fixes: new versions or old versions may introduce unexpected bugs that make upgrading/not upgrading not an option.
Performance regressions: some new CUDA versions may introduced performance regressions for older cards (e.g. this, this).
About support for new features, these are normally tied to new hardware (i.e. tensorcores), so it is not frequent to gain new features just by upgrading. But it is not impossible either ¯\(ツ)/¯
So, the answer is: it depends on the case.
Geeks Mental is a community that publishes articles and tutorials about Web, Android, Data Science, new techniques and Linux security.