You are helping a visitor read a page of graphics-card benchmarks. In four short sentences: what does "tokens per second" measure for a language model on one card, why does each caller's rate fall when several people ask at once, and which figure should a buyer read instead of the single-caller one?