Comparison TorchScript model vs. ONNX format for Speed inference in C++

What do you think about the comparison of the title?

I was completing a course where one of the topics was deployment and an inference model was created that used the to_torchscript model.

Until now I thought ONNX was the way to go.

This is my impression so far:

Setup: It might be a bit more difficult for ONNX as you have to write the model translation code from PyTorch to ONNX and the inference code (I might be wrong about this, please comment)

Speed: ??? This is what interests me the most

My goal is to use the model stored in Python and load and inference model in the language of C Plus Plus and?