Deciding at the edge

01Comparison

The same street, two architectures.

On the left, every camera streams to a distant data centre. On the right, an edge node on the block runs the model and sends only events. The sensors are the same; the decision moves closer and the backhaul carries a fraction of the bytes.

Figure as text: Cloud-only and edge inference on one block

A split view of the same city block: on the left, links from every pole run to a distant node; on the right, they run to an edge node on the block.

Cloud only

Every frame crosses the backhaul to a distant data centre.

Edge inference

An edge node on the block runs the model and forwards only events.

02Simulator

Time to decision.

Capture, transfer, inference: at the edge, that is the whole budget. Through the cloud, add the backhaul transfer and the round trip. Set the frame size, the links and the model’s inference time; every value is arithmetic on the inputs.

Figure as text: Decision latency at the edge and through the cloud

Two horizontal bars: the edge decision time in red, shorter, and the cloud decision time below it.

03Systems diagram

When the backhaul goes.

If the backhaul is cut, a cloud-only design stops deciding. An edge design keeps detecting and queues events for later delivery. That property is why edge inference is a resilience choice as much as a latency one.

Figure as text: Edge inference with a cloud for training and archive

A four-tier diagram: cameras, the access network, the edge node with the model, and the cloud and the operations centre.

• Cameras: Camera 1, Camera 2, Camera 3
• Access: Access
• Edge model: Edge model
• Operations: Operations, Training and archive

6 links. Scenarios: Normal operation; Backhaul cut.

04Timeline to 2030

Rules for models that decide.

The European Union’s AI Act entered into force in August 2024, with obligations for high-risk systems phased in over the following years³. A model at the edge that informs safety decisions will need its data, its versions and its performance on record.

Figure as text: Edge computing and AI governance

A timeline from 2016 to 2030 with edge computing standards, inference benchmarks and the AI Act.

1. 2016 Edge computing reference architecture (ETSI GS MEC 003). First edition of the multi-access edge computing framework.¹
2. 2019 Inference benchmarks published. Comparable measurements of inference on edge and data-centre systems.²
3. 2024-08 AI Act in force (Regulation (EU) 2024/1689). Phased obligations for providers and deployers of AI systems.³
4. 2027-08 Obligations for high-risk AI fully apply. The last of the Act’s phased application dates.³