Accelerator benchmark methodology

How inference chips are compared: suite choice, single-turn versus long-context multi-turn workloads, configuration parity such as multi-token prediction, and who runs the numbers.