Jetson inference gets slower after several minutes: power mode or thermal throttling?

Record the configured power mode and monitor frequencies, temperatures, utilization and memory during the same run. Use the tools documented for the installed Jetson Linux release. Correlate changes with per-frame latency rather than comparing unrelated snapshots.

Repeat with the intended cooling and enclosure at a known ambient temperature. Check input power adequacy and background workloads. A high-performance mode cannot compensate for insufficient cooling or a supply limitation, and it may change the conditions under which those limits appear.

Measure capture-to-result age and queue length alongside inference-call time. If call time stays steady while end-to-end delay grows, buffering deserves attention. Report stable sustained performance after warmup, with the hardware and power configuration included in the result.