Silicon LogiX
ITLet’s talk
Menu

04 / AI

Edge AI & TinyML

Measurable intelligence, in the real world

Local inference when memory, power, latency or connectivity require processing close to the sensor on MCUs or embedded platforms.

Platforms and tools follow your context. Some examples:

Python / NumPy / C++ / TinyML / Computer vision / Edge AI

AI

Edge AI & TinyML

Bring inference onto the device

  • Edge platforms

    MCUs · SoCs · acceleratorsSelected for available hardware

  • Models & conversion

    Formats · operators · runtimesTarget compatibility

  • Memory & model

    Quantisation · buffers · FlashActual device resources

  • Latency & power

    Timings · workloads · measurementsApplication constraints

  • Firmware integration

    Data · sensors · interfacesInference in the system workflow

  • On-target testing

    Accuracy · resources · performanceResults measured on the device

The technologies shown are examples.

Engineering support within your team, with shared tools and working methods.

Open graphic ↗

TECHNICAL SCOPE

Edge AI and TinyML on MCUs and embedded systems

Feasibility is checked on the actual platform: the model is only part of the cost because acquisition, buffers and preprocessing consume memory and time. Accuracy and resources are compared, considering quantisation and compatible formats. Behaviour remains defined for invalid data or insufficient confidence.

Activities and verification

  • Profile Flash, RAM, inference time and the complete processing chain.
  • Port and quantise while comparing results against the reference model.
  • Integrate tasks, sensors, communication and model version updates.

Initial decisions and integration

I define the platform, acquisition, runtime and resource budgets with firmware and model teams. I can support porting, profiling or product integration while retaining existing pipelines and tools when compatible with the target.

A verification example

Reference-model and exported-version results are compared on the same inputs. Memory, timing and behaviour with invalid data are measured on the device. Quantisation and optimisation are assessed alongside any changes in results.

Deliverables and acceptance criteria

Exported model, inference code, configuration and target measurements. Acceptance thresholds refer to documented data, frequency and operating conditions.

Where to start

Specify the MCU or SoC, available RAM and Flash, data source, maximum latency and energy requirements.

Can the model be updated separately from firmware?

It depends on runtime, memory and update architecture. Some products use a separate artifact; others distribute the model with firmware. The process is verified on the target.

Describe your context ↗

Insights

All articles ↗

THE NEXT STEP

A system ready
for its next step?

Let’s discuss your project