A Brain Too Big to Carry — On-Device vs Datacenter Inference

A robot can divide its computing between onboard reflexes and remote planning. SemiAnalysis explains that balance and motor control need frequent updates, while choosing the next task can allow a longer pause. Its example contrasts a control loop with a 10-millisecond deadline against a planner with 200 milliseconds per decision. That leaves very different room for a network round trip. The analysis argues that this separation makes some remote computation practical.

Predictable timing matters as much as average speed. Irregular network delays make it harder to know when the next instruction will arrive. The proposed division therefore keeps time-critical control nearby while allowing slower planning to use computing resources beyond the robot's power and hardware budget.