8B vs 14B sweet spot with TensorRT-LLM

[fastfield4]

A couple questions before I guess too hard. What model, what runtime, and roughly what context size?

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.