I’ve been having a lot of trouble getting tensorflow to run a gpu on windows 10. Using the CUDA setup I had to use tensorflow with a 1660 but ended up getting NaN values for loss with the 3080 (worked fine on 1660 and CPU).
I’ve tried using CUDA 11.0, 11.0, and a few different versions of cuDNN each of those, along with tensorflow nightly. With this setup I was getting CUDNN_STATUS_ALLOC_FAILED and CUBLAS_STATUS_ALLOC_FAILED in all of these cases. Again, this was on code that works fine on other GPUs and CPU.
Is it the case that the 3000 series cards are not officially supported yet?
I only opened a new thread because google wasn’t yielding a lot of results.