Tokens per second
How fast a model generates. Bound by memory bandwidth, not by arithmetic throughput.
Producing one token requires reading the weights that participate in it, so generation speed is roughly bytes-read-per-token divided by achievable memory bandwidth. This is why a card's bandwidth predicts local inference speed far better than its teraflops, and why an older card with fast memory can beat a newer one with more compute.
It is also why quantizing a model speeds up generation — fewer bytes to read — while leaving prompt processing alone or slightly slower.