Hi,
Both TX2 or TX2i are good for single 4K video decoding. And for doing deep learning inference, please try DeepStream SDK.
We have implementation for IP camera source. you can try to set source to type=2(URI) or type=4(RTSP). Please refer to document:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_ref_app_deepstream.html#source-group