Workload planning desk / Reading notes
Benchmark the work you intend to run
A comparison method that survives changes in hardware labels.
Freeze the workload
Record input dimensions, precision, algorithm and software versions. Different numerical settings can change both speed and output quality.
Separate warm-up
Report initialization and steady-state execution separately. Include repeated measurements and variation when the difference affects a resource decision.
Include end-to-end time
Account for staging, transfers and result writes. A faster kernel does not guarantee a faster completed job if data movement dominates.