There is no answer that is always right. There is a right choice for each case, and often it is a mixed architecture.

The four criteria

• Latency: if the answer must arrive in under a hundred milliseconds, the edge is almost mandatory.
• Bandwidth: streaming video continuously from many cameras is expensive; sending only events is not.
• Privacy: images and audio processed on site never leave the company.
• Cost over time: edge hardware is paid for once, cloud compute every month.

The hybrid architecture

The most common solution: inference at the edge, training and historical analysis in the cloud. Devices send events and meaningful samples, the cloud retrains models and redistributes them. You get the responsiveness of the edge with the improvement capacity of the cloud.

Mistakes to avoid

Choosing hardware before measuring the real workload, forgetting remote model updates and underestimating environmental conditions: dust, heat and vibration put any board to the test.