How can an SBC performance comparison avoid rewarding only short turbo bursts?

Use the same meaningful task, input and output-quality requirement rather than comparing unrelated demonstration scores. Record software versions, architecture, thread count and acceleration paths. A board using a hardware codec is not performing the same CPU workload as one decoding in software.

Run long enough to observe stable temperature and operating frequency with the intended cooling. Measure power at a consistent boundary, such as the complete board input. Include startup and idle behavior if those occupy much of the real deployment cycle.

Report sustained throughput, tail latency and energy per completed task. Check that specialized hardware support is available in the maintained image you plan to deploy. The fastest short result is less useful if its software configuration cannot support the required peripherals or long-term workload.