Custom benchmark runs can useful for debugging and testing purposes. Please note that custom runs do not produce an overall benchmark score.
From the Home screen, tap on the Customize button.
The initial settings shown on the Customize screen are the default for standard benchmark runs.
If Automatic is selected, NNAPI will pick the best accelerator for each task.
If you select another accelerator, NNAPI will try to run the task on that specific accelerator.
Select AI Models
You can pick and choose which models to include in the run. You must pick at least one.
Allow F16 precision for F32
Most accelerators do not support F32. Enabling this option reduces the precision from F32 values to F16 when performing the calculations required for the task.
Enabling this setting will run the custom benchmark for 10,000 rounds for each selected model. Please note that the benchmark only runs on NNAPI with this setting. The benchmark will not run on the CPU or GPU.