The Two Clocks (TTFT, TPOT, and why goodput is the number)

One request laid on a ruler, bracket by bracket, every millisecond computed from the model and the card. Below it, the thing that owns both brackets and its two dials. Below that, ten requests in one second, and the number the run is actually scored on.

Before you start. Inference Engineering (serving an LLM is a systems problem) walked the whole trip a request takes through a serving system and pointed out that a reply hides two different waits inside it. This lesson puts real numbers on both waits and shows the one dial that trades one against the other.