Field Guide · concept

Also known as: on-device AI, edge inference

Edge AI is the practice of running machine-learning models directly on local devices — single-board computers, cameras, sensors — instead of sending data to the cloud for inference.1

sensor cloud server result network round trip · latency sensor accelerator on-device result local · instant cloud edge
Cloud inference ships raw data to a distant server and waits for the round trip; edge AI runs the model on an accelerator right next to the sensor, returning a result immediately and sending only the small answer onward — the source of its latency, privacy, and offline advantages.

Overview

The appeal is concrete: inference next to the data is lower-latency, keeps private data on the device, and keeps working without a network. A doorbell that recognises a face locally answers in milliseconds and never uploads the video; a cloud version has to survive a round trip and trust a remote server with the footage. Bandwidth costs drop too, because only the small result travels rather than a continuous raw stream.

The cost is that small devices have limited compute, so edge AI leans on efficient, quantised models and on dedicated AI accelerators. Rather than run a full-precision network on a general CPU, edge deployments shrink the model — lower numeric precision, pruning, distillation — and hand the heavy math to purpose-built silicon like Google Coral’s Edge TPU or the GPU on an NVIDIA Jetson. It is a specific case of the broader move toward edge computing.

Edge vs cloud inference

  Cloud inference Edge AI
Latency Network round trip Immediate, local
Privacy Data leaves the device Data stays put
Offline Fails without a link Keeps working
Compute available Effectively unlimited Small, power-bound
Model size Large, full precision Quantised, compact
Ongoing cost Bandwidth + server time One-time hardware

Where it fits

Edge AI shows up wherever round-tripping to a server is too slow, too costly, or impossible: factory cameras, doorbells, robots, and home automation hubs. In a GopherTrunk-style deployment, edge AI could classify or flag activity in decoded data on the same node that does the radio work — spotting a signal of interest, sorting traffic, or triggering a recording — and send up only the results rather than every sample, keeping a remote capture node useful even on a thin link.

Sources

  1. Edge computing — Wikipedia, on processing data near its source, including on-device inference. 

See also