I just replaced an old GeForce GT1030 card with a new GeForce RTX 3060 card. I had StarNet2 2.0.0-34 with the CUDA modules running in PixInsight 1.8.9-1 (Win10 x64) with the old card (replace tensorflow dll, install CUDA 10.1 and copy the CUDNN \bin and \lib files) and it was working (100% on the GPU in Starnet2).
After installing the new card, I installed the latest version of the nVidia driver (516.59) and checked that the same tensorflow, CUDA and CUDNN files were still there.
I ran Starnet2 and it took a few minutes before it did anything (busy cursor) then flew through the process but generated junk. See attachment.
I tried uninstalling and reinstalling CUDA 10.1 but that didn’t help.
It was using a bunch of GPU memory but very little GPU processor time.
I managed to get it working. I found a copy of tensorflow.dll version 2.8.0 for GPUs. I used that with version 11.2 of CUDA and 8.1.1 of CUDNN. Maybe it was because my nVidia driver is the latest version (516.59). I was hesitant to downgrade the driver and I didn't have to with these versions of tensorflow, CUDA and CUDNN. When I run nvidia-smi it says I have CUDA version 11.7. Maybe that's the version in the driver???
I don't understand why it's so hard to find a compiled version of tensorflow.dll. I still haven't found a compiled version of 2.9.0.
I spent all day trying different versions of tensorflow, CUDA and CUDNN before I finally found a set that worked. Kind of frustrating. The whole thing seems totally arcane to me.
Edited by MAT_Blue, 17 July 2022 - 03:55 PM.