Edge AI in production: measure the system, not just the model
Latency, energy, acquisition and updates: checks that matter when AI runs close to industrial machines.
Running a model near the machine collecting data can reduce connection dependency and make response times more controllable. This is the idea behind on-device, or edge, AI. The decision involves more than whether the model fits in memory. An industrial application must run for hours, handle errors and updates, and maintain performance compatible with the process.
The complete information path
Consider hypothetical visual inspection on a conveyor. A camera captures an image, the system prepares it, the model produces a result and the application associates it with the correct part. Another component then makes the process decision. Measuring only model time excludes transfers, queues and synchronization. A fast prediction can still arrive too late if the rest of the path is slow.
Distinguish latency—the time for one processing operation—from the capacity to sustain incoming work. A system may handle one image well but accumulate delays when the conveyor accelerates. Testing should reproduce plausible input frequency, data sizes and load. If an image is lost, software must detect the event and apply a defined procedure.
Read accuracy, time and energy together
MLCommons' MLPerf Tiny provides shared measurements for resource-constrained systems, considering accuracy, latency and energy. It is a methodological reference: its workloads do not automatically represent an industrial line or a large language model. Benchmarks help frame comparisons, but the complete system must be checked on its intended task.
Lower numerical precision or a simpler model may improve execution on limited hardware. The effect on quality cannot be assumed. If errors increase on a rare but important defect, average speed improvements may not help. Compare versions on identical inputs and preserve examples where decisions differ.
Longer trials reveal different problems
A short demonstration does not necessarily show steady-state behavior. Temperature, resource saturation and queue growth can change performance. Observe the application over a representative period, including starts, pauses and restarts. Energy consumption also affects cooling and enclosure choices, not merely electricity costs.
Real inputs change with lighting, optics, vibration and materials. A model working on a laboratory camera may need new validation after acquisition changes. Versioning the model without recording camera configuration leaves the history incomplete. Diagnosing errors requires knowing which combination of components and settings was actually operating.
Local does not mean free from every risk
Keeping processing on site can limit some transfers but does not itself guarantee privacy or security. Logs, diagnostic copies and update services may still transmit data. Specify what stays on the device, what is sent elsewhere and why. Physical access, credentials and retention of collected material also need consideration.
Updates need a procedure for identifying the installed version and recovering the previous one when necessary. Progressive deployment to a limited set of devices can reduce the impact of errors. If connectivity fails, it must be clear what continues working and which services become unavailable. A hybrid architecture can keep essential functions local and use remote resources for deferred analysis.
An infrastructure question for EL-AI
EL-AI regards industrial and collaborative robotics as a direction to explore. Studying where perception and analysis run is therefore relevant, but does not imply the company already has an edge platform or production installations to showcase. Start by describing the physical task and its constraints, linked to the discussion of a first collaborative robotics project.
For initial assessment, prepare a brief covering input flow, latest useful response time, quality criteria, error behavior and update procedures. Compare platforms afterward. This sequence avoids choosing on a single hardware specification and focuses on what makes the application sustainable in its actual environment.
Article prepared with AI assistance and verification of the cited sources. Application examples are hypothetical unless stated otherwise. Sources consulted on September 20, 2026.
Illustrative AI-generated cover; it does not depict actual EL-AI people, premises or installations.