What is the GPU that is being recommended for inference and fine tuning? Is a 4090 sufficient

[IsaacSinn]

[IliaLarchenko]

It depends on your setup, e.g., the number of cameras and their resolution and FPS, but I think 4090 will be more than enough for many cases.

I ran inference on MacBook Pro M1 without GPU using feed from 3 cameras, 640x480 - and it did ~25 predictions per second (with cameras FPS 30). It will be much faster on 4090, so it is more than sufficient for inference.

4090 is also sufficient for training. I managed to train the policy even on a 2080Ti in a reasonable time. The model itself is very small, so it will fit any Nvidia GPU. Loading data (especially images) was the main bottleneck during the training.