Every chip that runs a language model is sold on three numbers: memory capacity, memory bandwidth and compute. This field guide shows which part of the work each one controls, with seven real chips from a Mac Studio to an NVIDIA B200.
In the middle there's a calculator. Pick a model, a precision and how long the prompt and the answer are, and it shows how long each chip spends reading the prompt and writing the answer, and how it copes with many users at once.