Blog · Industry Updates
Edge AI Technology Update
Onboard inference became the default architecture in 2026, and the number that sells it — TOPS — is the number buyers most often misread.
Overview
- TOPS is theoretical peak compute. It does not tell you frame rate, latency, or whether your model fits. Two modules quoting the same TOPS can differ by a wide margin in delivered accuracy after quantisation.
- Published engineering budgets for 2026: a 6 TOPS module runs a small detection model at 1080p around 60 Hz with 20–30 ms of inference latency, landing near 40–50 ms end to end. A 1 TOPS module reaches real time only by dropping frame rate, resolution or model depth.
- Weight and power are the real constraints. Typical figures are 40–70 g and 3–5 W at 1 TOPS, 80–120 g and 5–9 W at 6 TOPS — and sustained load, not peak, is what causes thermal throttling mid-mission.
- Power quality matters as much as draw: NPU current comes in bursts that can reach two to three times average. A power system that cannot supply the burst will brown out the module under sustained load.
Collected 9 October 2026 from public module vendors and integration guidance, including a 12 TOPS sub-10 W companion computer announced 30 September 2026. Montline publishes compute figures only with the model, quantisation and measured latency attached.
Related pages
- Procurement and Selection GuidesView
- Company ProfileView
- Technical StrengthView
- Industrial MultirotorView
More in this area
- Civil Drone Industry Applications UpdateRead
- Drone Payload Technology UpdateRead
- Battery and Power Technology UpdateRead
- VTOL Civil Application UpdateRead
- Observations from Drone Logistics PilotsRead
- Drone Surveying Technology UpdateRead
- Thermal Imaging Application Technology UpdateRead
- Edge AI Technology UpdateThis page
- Drone and Enterprise Software Integration NewsRead
- European and US Civil Drone Market UpdateRead