CUDA Toolkit and Minimum Required Driver Version for CUDA Minor Version Compatibility - Documentation issue

Up to the documentation of CUDA toolkit 12.2.2 (Release Notes :: CUDA Toolkit Documentation) we had the following requirements:

Documentation of toolkit 11.3.0 (Release Notes :: CUDA Toolkit Documentation) says:


and 11.3.1 ( Release Notes :: CUDA Toolkit Documentation (nvidia.com)):

whereas the latest documentation ( CUDA 12.0 Release Notes (nvidia.com)) says:

I’m concerned about the lowering of the requirements for toolkits 11.2.* that went from minimum driver versions 460.27/32.03 in documentation 11.2.2 down to 450.80.02 in documentation 11.3.0 and after.

We have V100 and A100 cards with Driver Version: 455.23.05. Using toolkits up to version 11.1 works fine. Any attempt to use a toolkit from 11.2 or above fails with cuda error 222 (cudaErrorUnsupportedPtxVersion), what makes me believe that the actual documentation is misleading but was correct up to version 11.2.2 (Release Notes :: CUDA Toolkit Documentation).

Note that this applies to the case where you want to generate PTX (e.g. with -arch compute_80 -code=compute_80 for an A100 card) to be JIT compiled. Generating binaries (e.g. with -arch compute_80 -code sm_80 for an A100) works fine with both 11.1 and 11.2 toolkits.

R455 is EOL according to
https://docs.nvidia.com/deploy/cuda-compatibility/index.html#use-the-right-compat-package

Yes, I saw this one. However there is no such deprecation notice in the release note of the 11.3.0 toolkit. And, if I understand the version numbering correctly, it’s a bit awkward to step back to a lower version than indicated at release time.