Compute Tier Placement
Deciding which of on-device, on-premises, metro, regional or central compute a workload belongs to, from its latency, data volume and autonomy requirements.
"Edge" spans several orders of magnitude of distance and capability, and the tiers have genuinely different properties.
On-device has no network dependency at all and the least compute; it is the only tier that continues working when the link is down, which is the deciding factor for anything safety-related. On-premises — a gateway or small cluster on site — has more capability and serves the whole local installation, which is where data reduction usually belongs. Metro and regional provider edges give tens of milliseconds and elastic capacity without local hardware. Central has unlimited capability and the most latency.
Three questions place a workload. What latency does the control loop require — a millisecond response cannot leave the device. What data volume is involved — a camera producing gigabytes per hour must be reduced locally, because the network cost dominates everything else. And what must keep working when disconnected — autonomy is the requirement that most often forces work down a tier and is most often discovered late.
The common error is treating it as one tier and putting everything at the provider edge, which gives neither the autonomy of local compute nor the economics of central.