AI Glossary / MLOps and deployment / Intermediate
Edge computing
Definition
Edge computing places some data processing close to the devices, people or physical processes that produce or use the data. The processing may run on a device, a nearby gateway or a local server. Its purpose is to meet practical needs such as responsive operation, limited connectivity or reduced transmission of raw data.
What it means
The “edge” is relative to a network or application architecture, not one specific device type. A camera can analyze images locally, while a nearby site server combines results from several cameras and a central cloud service manages models. These locations can cooperate. Edge computing is broader than edge AI: local filtering, storage and control count as edge computing even when no machine-learning model is involved. Definitions overlap with related terms, so describe where each operation actually runs.
How it works
A system assigns tasks according to delay, bandwidth, computing capacity, reliability and data-handling requirements. Local components may preprocess measurements, run inference, cache information or make bounded decisions. Central services may provide larger-scale analysis, model training and fleet management. Synchronization and updates connect the layers. When the network fails, the system needs a defined local behavior, and when it recovers, delayed data or decisions may need reconciliation. Moving computation closer changes these responsibilities rather than making them disappear.
Three practical examples
- A factory camera checks a product locally and sends selected events to a central dashboard instead of streaming every frame continuously.
- A mobile application performs a small recognition task on the phone so that its core feature can work during poor connectivity.
- A remote monitoring gateway filters repeated readings and stores a temporary buffer, forwarding summaries and important events when a connection is available.
Think of it this way
Think of a branch office handling routine requests locally while head office coordinates policy and larger projects. Local decisions can be quicker, but the branches need current instructions, reliable equipment and a way to resolve inconsistencies. The analogy does not imply that edge devices exercise independent human judgment or that every task should be decentralized.
Worked example
Suppose a hypothetical inspection system needs a result within 100 milliseconds. Sending a frame to a remote service takes 45 milliseconds, queueing and inference take 35, and returning the response takes 30. The total is 110 milliseconds, so that path misses the target even though inference itself is fast. A local path with 12 milliseconds of preprocessing, 40 of inference and 8 of output handling totals 60 milliseconds. These illustrative numbers show why end-to-end delay matters. They do not establish a real service guarantee: test variability, peak load, thermal throttling and failure conditions before accepting the design.
Illustrative case study
Fictional case: Fir Retail pilots local shelf-image analysis in stores with uneven connectivity. The team keeps the core detection step on a store device and sends event summaries centrally. It discovers that model updates must be staged carefully because several device generations have different memory limits. It adds version reporting and a rollback procedure, then tests operation during a network outage. The case illustrates deployment management rather than a claim about sales, accuracy or customer outcomes.
Common misconceptions and mistakes
Do not assume that edge processing is always faster or cheaper: a small device may be slower than a powerful remote server. Local processing does not automatically guarantee privacy; raw data, derived results, logs and update channels still require appropriate handling. Edge systems may remain cloud-connected and centrally managed. Also, installing a model on a device is only part of the work. Monitoring, physical maintenance, patching and reliable recovery can dominate a large fleet's operational effort.
Business use
Businesses should begin with a concrete requirement: an acceptable response time, a bandwidth constraint or an offline operating need. Compare architectures against that requirement using realistic loads and devices. Include purchase, energy, maintenance and update costs. For AI, evaluate model quality after any size or precision reductions needed for local hardware. Define who owns the device fleet and how degraded operation is communicated to people relying on its results.
People also ask
Does edge computing require AI?
No. Many local processing tasks use ordinary rules or signal processing.
Does it eliminate the cloud?
No. Hybrid designs often use both local and central computing.
Is a nearby gateway considered the edge?
It can be, depending on the architecture. State its location and responsibilities rather than relying on the label alone.
Check your understanding
A team reports a model's local inference time as 20 milliseconds and promises a 20-millisecond response. What is missing?
Answer: The rest of the path, including data capture, preprocessing, queueing, communication where applicable and output handling, plus variation under realistic operating conditions.
Related terms
- Cloud computing
- Latency
- edge AI
- fog computing
- gateway
- on-device inference
- fleet management
Learn next
Draw the complete path from data capture to the action or display that uses the result. Mark where each stage runs and measure its delay. Then repeat the exercise for a network outage and a failed update, because those scenarios reveal requirements that a simple average-latency comparison misses.
More in MLOps and deployment
- Discrete cosine transform - The discrete cosine transform, or DCT, represents a finite sequence of real numbers as weights of cosine patterns.
- Dyadic distribution - In source coding, a dyadic distribution is a discrete probability distribution whose positive probabilities are integer powers of one half, such as 1/2, 1/4...
- Dynamic Markov compression - Dynamic Markov compression, or DMC, is a lossless compression approach that adaptively models the probabilities of incoming bits using a changing state...
- Elias delta coding - Elias delta coding is a self-delimiting binary code for positive integers.
- Elias gamma coding - Elias gamma coding is a self-delimiting binary representation of positive integers.
- Elias omega coding - Elias omega coding is a self-delimiting binary code for positive integers that recursively represents the lengths needed to describe a number.
Sources
Primary references: NIST, Formal Definition of Edge Computing, https://www.nist.gov/publications/formal-definition-edge-computing-emphasis-mobile-cloud-and-iot-composition ; NIST SP 500-325, https://csrc.nist.gov/pubs/sp/500/325/final . Institutional publication search excerpts consulted; full architecture-specific and editorial review remain pending.
Previous: Edge case Next: Edge data integration