Embedded AI

Artificial intelligence right on the module – fast, secure, no cloud.

Our byteENGINE System on Modules with integrated NPU run neural networks locally. And we help you build and optimize your AI model so that it runs reliably on your target platform.

Discuss your AI project View modules with NPU

  • up to 2.3 TOPSNPU performance on the module
  • 3 platformsNXP i.MX 8M Plus, i.MX 93, ST STM32MP25
  • Swiss madeDevelopment & support from Winterthur

Why edge AI?

When inference happens directly on the device instead of in the cloud, industrial applications gain tangible benefits.

Real time

Decisions in milliseconds – with no network latency. Essential for quality inspection, control and safety.

Data privacy

Images and sensor data never leave the device. Ideal for medical technology, public spaces and sensitive production data.

Works offline

Runs without an internet connection – in the field, in vehicles, in agriculture or in an isolated OT network.

Efficient

The NPU is many times more efficient than the CPU. Fanless designs at just a few watts and with no ongoing cloud costs.

AI on the embedded NPU

A neural processing unit (NPU) is a dedicated accelerator for neural networks. It offloads the CPU, so your application, communication and real-time tasks keep running in parallel. These byteENGINE modules come with an NPU:

2.3TOPS

NXP i.MX 8M Plus · 4× Cortex-A53 + Cortex-M7

byteENGINE IMX8M Plus

  • Most powerful NPU in our portfolio
  • Integrated image signal processor (ISP) for camera applications
  • Ideal for machine vision, object detection, multiple models in parallel
  • Toolchain: NXP eIQ, TensorFlow Lite, ONNX Runtime

Vision & multi-model

1.35TOPS

ST STM32MP257 · 2× Cortex-A35 + Cortex-M33

byteENGINE STM32MP25 OSM

  • Compact OSM form factor (30 × 30 mm)
  • Two camera inputs, Full HD display, 3D GPU
  • Ideal for vision and anomaly detection in industrial applications
  • Toolchain: ST Edge AI Suite, X-LINUX-AI, TensorFlow Lite, ONNX

Compact & industrial

0.5TOPS

NXP i.MX 93 · 2× Cortex-A55 + Cortex-M33

byteENGINE IMX93 OSM

  • Arm Ethos-U65 microNPU – highly energy-efficient
  • EdgeLock Secure Enclave for secure devices
  • Ideal for sensor data analysis, keyword spotting, simple vision
  • Toolchain: NXP eIQ, Arm Vela Compiler, TensorFlow Lite

Low power

No NPU needed? Smaller models can also run on the CPU or GPU of our modules without an NPU (e.g. the TI AM62x family) – we will advise you on the right platform for your application.

Typical applications

Embedded AI fits wherever devices need to detect, classify or predict on their own.

Application areaExampleModel type
Industry & automationOptical quality inspection, defect detection on components, predictive maintenance based on vibration or current dataClassification, anomaly detection
Smart city & infrastructurePeople and vehicle counting, occupancy detection – privacy-compliant, directly on the deviceObject detection
Medical technologySignal classification, image pre-processing, assistance functions without data transmissionClassification, segmentation
Transportation & agricultureDetection of objects, plants or animals in the field, even without network coverageObject detection
Test & measurementIntelligent evaluation of measurement signals, automatic classification of test resultsTime-series models

From model to target platform – we support you

A model that works well on a PC with a GPU will not automatically run efficiently on an embedded NPU. Supported operators, quantization, memory footprint and latency all have to match the hardware. That is exactly where we come in.

  1. FeasibilityAnalysis of your use case and your requirements for accuracy, latency and power. Selection of the right platform.
  2. Model creationSelection of a suitable architecture (e.g. from model zoos) and training or fine-tuning with your data.
  3. OptimizationINT8 quantization, pruning, adaptation of unsupported operators – so that as much as possible runs on the NPU.
  4. ConversionCompilation for the target NPU with the vendor tools (NXP eIQ, Arm Vela, ST Edge AI).
  5. BenchmarkingMeasurement of inference time, accuracy and memory on real byteENGINE hardware.
  6. IntegrationIntegration into your Linux system and your application – including camera pipeline and updates.

What you get from us

  • Platform recommendation based on measurements instead of datasheet figures
  • Optimized model, quantized and compiled for your target NPU
  • Yocto/Linux integration of the AI runtime into your product image
  • Long-term support – from prototype to series production and beyond

Already have a model? We port and optimize it for your target platform. Just getting started? We accompany you from the initial idea to a working prototype.

Software stack on the byteENGINE

  • Your application C/C++, Python, Qt
  • Vision pipeline GStreamer, OpenCV
  • AI runtime TensorFlow Lite, ONNX Runtime
  • NPU drivers & delegates eIQ, Ethos-U, X-LINUX-AI
  • Embedded Linux Yocto / Debian, meta-bytesatwork
  • Hardware byteENGINE SOM with NPU

Frequently asked questions

Which model formats are supported?

Primarily TensorFlow Lite and ONNX. Models from PyTorch or Keras can usually be converted. Which operators run directly on the NPU depends on the platform – we check this for you.

What does TOPS mean – and is more always better?

TOPS (tera operations per second) describes the theoretical computing power. What really matters for your application, however, is the actual inference time of your model, the NPU's operator coverage and the power consumption. That is why we measure on real hardware.

Can I test the NPU in advance?

Yes – with our byteDEVKITs you can evaluate the modules and the AI examples right away. Just get in touch.

What about security and the Cyber Resilience Act?

AI functions are part of your product and therefore also part of your CRA obligations. We integrate models and runtimes so that they work together with secure boot, update mechanisms and our lifecycle management.

Bring AI to your device

Tell us about your application. We will assess the feasibility and show you which platform and which model fit your requirements.

Get in touch Modules & DEVKITs

Datenschutzhinweis

Diese Webseite nutzt externe Komponenten, wie z.B. Google Analytics, Google Maps, und Youtube, welche dazu genutzt werden können, Daten über Ihr Verhalten zu sammeln. Datenschutzinformationen

Notwendige Cookies werden immer geladen