Inference time covers only the interval from inference start to end. Full device response time uses the events agreed for the site, such as an external trigger and receipt of that workpiece result at the controller. Acquisition, transfer, queuing, preprocessing, postprocessing and result delivery can all affect response time.
Check an individual timing record
Use one workpiece identifier to match events. Check whether inference timing includes data copies or waiting for execution to finish. Record device response time and inference time separately under the test conditions. Keep records of missing results, timeouts and retries.
Confirm machine cycle time
Measure machine cycle time using consecutive completion events on the actual line. Concurrent tasks and queues can change the interval between completions. Record equipment, software, image settings and parallel tasks. State which samples the statistics cover.
If the application retains only inference logs, check whether workpiece identifiers can link them to external trigger and controller records. Without matching records, report the measured inference interval and mark full response time as pending measurement.
Related service: Vision algorithms and model engineering.