I just spent an hour fighting with Starnet to use my GPU without luck until I tried something; I have not seen this documented elsewhere so I wanted to share in the event it helps someone. I was following this guide to get my GPU (RTX 3090) working with Starnet. Problem was, it just did not work. It'd always just hang when I tried to start it.
I tried removing/reinstalling CUDA, same goes for cudnn. I tried both 10.1 and 11. Everything just did not work. In the logs, I could see this as the last entry:
tensorflow core adding visible gpu devices 0
I realized then that the issue may be Tensorflow. I swapped out to TF 2.5 (instead of 2.3 which most posts tell you to use) and then realized it requires CUDA and cudnn 11. Once all of that was in place, everything worked like a charm.
tl;dr:
If you have a newer GPU and Starnet GPU doesn't work based on the common guides you see, swap out for TensorFlow 2.5 and the CUDA / cudnn versions it likes (11.4).
I tried removing/reinstalling CUDA, same goes for cudnn. I tried both 10.1 and 11. Everything just did not work. In the logs, I could see this as the last entry:
tensorflow core adding visible gpu devices 0
I realized then that the issue may be Tensorflow. I swapped out to TF 2.5 (instead of 2.3 which most posts tell you to use) and then realized it requires CUDA and cudnn 11. Once all of that was in place, everything worked like a charm.
tl;dr:
If you have a newer GPU and Starnet GPU doesn't work based on the common guides you see, swap out for TensorFlow 2.5 and the CUDA / cudnn versions it likes (11.4).