ASSAY-003
Sustained Jetson Inference Under a 15 W Power Envelope
Characterizing sustained vision-model throughput, thermals, and clock behavior on a Jetson Orin Nano held at a strict 15 W budget for one hour.
- Status
- COMPLETE
- Domain
- EDGE · COMPUTE
- Date
- 10 AUG 2026
- Claim
- Sustained INT8 throughput at 15 W settles 18% below the 60-second benchmark figure once thermal steady state is reached.
- Configuration
- Jetson Orin Nano 8 GB, 15 W mode, MAXN clocks disabled, stock heatsink, 22 °C ambient.
- System Boundary
- Module power measured at the DC barrel input; fan excluded via separate supply.
- Provenance
- All raw logs captured with tegrastats at 1 Hz and archived alongside the notebook.
- Software Rev
- JetPack 6.1, TensorRT 10.3, test harness commit a1b2c3d.
- Environment
- Closed office, no forced airflow beyond the module fan.
- Reproducibility
- Run scripts and parts list in the repository; one-hour soak required before recording.
Question
What does a Jetson Orin Nano actually sustain - throughput, clocks, and
temperature - when held at a strict 15 W budget for a full hour?
Method
A YOLOv8n INT8 TensorRT engine runs continuously on looped 1080p video for
60 minutes. Power is measured at the DC input; tegrastats logs clocks and
temperatures at 1 Hz. The first 10 minutes are discarded as warm-up.
POWER14.9 W avg
FPS (60 s)112
FPS (steady)92
T(GPU) steady71.3 °C
Results
Sixty-second benchmark throughput was 112 FPS. After thermal steady state
(reached at ~22 minutes), sustained throughput settled at 92 FPS - 18% below
the benchmark figure - with GPU clocks stepping down twice as the heatsink
saturated.
Reproducibility
Run scripts, engine build flags, and raw tegrastats logs are archived with
the notebook. One-hour soak at 22 °C ambient required before recording.
