recently, I’ve been forced to perform a manual downgrade to CUDA v. 11.2.0. and CudNN v. 8.7.0 for the sake of running TensorFlow on Windows and WSL2. On the other hand, I am still using the newest game-ready drivers (527.68 for NVIDIA GeForce RTX 3060 for Notebooks).
I wonder, will I encounter any inconvenience by mixing newer drivers with older versions of CUDA and CudNN? In theory, it shouldn’t be a problem, as newer NVIDIA drivers are back-compatible with older CUDA versions - on the other hand, I want to make sure that it will not cause any problems in the future.
Thank you for your support!
If I understand Minor Version Compatibility correctly, there is a minimum driver requirement that I must satisfy for each CUDA Toolkit (and this is generally the driver version that the designated CUDA Toolkit was released with) but there should not be any problem with using older CUDA Toolkit and newer drivers. So, for CUDA v. 11.2 the minimum drivers are >= 450.80.02* and >=452.39* (for Linux x86_64 and Windows respectively), but newer drivers are also completely fine?
If I understand that correctly, I am gonna marked this as an answer and close the thread. I think that it may be helpful for some people to come, as this is not explicitly stated in the FAQ (it makes sense logically, but it can be more tricky with drivers really). Generally, the matter is not trivial, as newer releases of TensorFlow do not run on CUDA 11.8, so users must downgrade to 11.2. On the other hand - for various reasons - you may not want to stay on the older drivers just for the sake of the CUDA.