EVT / DVT checklist for AI hardware
AI products fail validation for boring reasons: a throttle looks like bad inference, or a camera flex cracks after ten thermal cycles.
EVT (engineering validation test) proves the integrated system works on bench and in targeted environmental tests. DVT (design validation test) proves the design meets requirements across units, orientations, and stress. For AI hardware, both stages must include thermal-soak with inference load, camera calibration drift, and OTA rollback, not just electrical continuity.
EVT gates
- Power and sequencing under GPU/NPU peak load
- Thermal steady state in worst-case enclosure orientation
- Vision pipeline latency and dropped frame budget
- RF pre-scan for antenna keepouts and cable routing
- Model version pinned in firmware with rollback path
DVT gates
- Multi-unit spread on camera color and exposure
- Vibration and drop per product class
- Manufacturing test coverage matches field failure modes
- BOM alternates qualified without re-tuning models
Next step: NPI + scale and Jetson carrier manufacturing.
EVT / DVT FAQ
How many EVT units do we need?
Often 5–20 for bench, plus environmental samples. AI vision products may need more for camera lot variation.
Should we run ML accuracy tests in DVT?
Yes, on production-representative hardware with frozen datasets. Treat accuracy regression like any other spec.