Edge AI

Your data. Your model. Your edge.

From saving bandwidth and energy to more responsive real-time performance, implementing AI in your embedded applications offers massive benefits beyond the buzzwords. Nordic Semiconductor offers two unique technologies, Neuton models and Axon NPU, exclusively to our customers, to cover the industry's broadest range of devices, applications, and customer needs.

 Neuton models

  

Axon NPU
Custom Neuton models are ultra-tiny edge AI models built from your data using our patented network-growing algorithm, ideal for running edge AI on any Nordic SoC or SiP using its main application core (CPU).    The Axon NPU is our dedicated AI accelerator core, designed to increase the speed and efficiency of TensorFlow Lite models, built into our most capable SoCs. 
 Neuton_Logo_horizontal_Gradient_RGB.png    Axon_Logo_horizontal_Gradient_RGB copy.png
     
     

10x

 10x

< 5 KB 

  

15x 

8x

7x 
Smaller memory footprint than TensorFlow Lite models Faster and more energy efficicent than running TensorFlow Lite models on the CPU  Average memory footprint of custom Neuton models created with our framework    Faster and more energy efficient than running the same TensorFlow Lite model on the CPU  More energy efficient compared to the closest competing product Faster inference compared to the closest competing product
           

 

 

  

Nordic Edge AI Lab

Nordic Edge AI Lab is your gateway to ultra-efficient edge intelligence. In the Edge AI Lab, you can create high-performance AI models ready to deploy on Nordic’s ultra-low-power SoCs, for both CPU-run and NPU-accelerated edge AI. 

Choose between three different paths for edge AI model creation: 

  • Custom Neuton models - Ultra-efficient CPU-run edge AI for any Nordic SoC

  • Text-to-wake-word - No-data generation of wake word models for Axon NPU

  • Model builder for Axon NPU - Build models with your custom dataset and architecture


 Documentation for the Edge AI Lab is available here.

Go to the Nordic Edge AI Lab

    Development paths

    Expand for more details

  • Custom Neuton models

    To create a custom Neuton model for your application there is no data science knowledge or experience required, only a good dataset and a can-do attitude, the Edge AI Lab handles the rest.Development of models are completed in three easy steps: 

     

    1. Upload your dataset as a labeled CSV and select the target column
    2. Select signal processing and feature extraction settings - or use automatic selection
    3. Wait for training to complete and download the model as a compiled C library 

    Neuton models are grown neuron by neuron using our patented algorithm, which creates the most accurate, efficient and smallest network possible. This eliminates the need for manual optimization and user input of network architecture, creating unique neural networks with no unnecessary neurons or connections, ensuring a ulta-tiny memory footprint and lightning-fast execution. 

     

     

  • Text-to-wake-word

    Create custom wake-word detection models for the Axon NPU without collecting any audio data. Simply provide your desired wake phrase as text and the tool generates a deployment-ready wake-word model, synthesizing all necessary training data internally.

     

    1. Enter your desired wake-word or phrase as text
    2. The tool synthesizes and augments the necessary audio training data
    3. A wake-word model is automatically trained and optimized
    4. Download the finished model, compiled for inference on Axon NPU

     

    Text-to-wake-word gives you a fast-track to voice activation, with no AI expertise or data collection required.

  • Model builder for Axon NPU

    Design and train fully custom classification or regression models based on TensorFlow, using your own datasets and neural network architectures. The resulting models are fully compatible and compiled for the Axon NPU, ready to build into your nRF Connect SDK application. This is the maximum-flexibility path for developers who need full control over model design and training.

     

    1. Upload your labeled dataset and select target column
    2. Specify pre-processing and feature extraction
    3. Specify your custom architecture
    4. Train the model and download a model compiled for Axon NPU

     

    Full control over data, architecture, and task, to take advantage of dedicated NPU acceleration for your application.