Feedback
Feedback

If you are having issues with the exercises, please create a ticket on DevZone: devzone.nordicsemi.com
Drag & Drop Files, Choose Files to Upload You can upload up to 2 files.

Axon NPU

The Axon NPU (Neural Processing Unit) is a dedicated AI accelerator core, designed to increase the speed and efficiency of TensorFlow Lite models.

Although TensorFlow Lite models can be run on the nRF54L Series using the CPU, the nRF54LM20B SoC is equipped with the Axon NPU to greater efficiency. Running these models on the Axon NPU makes them up to 15x faster and up to 10x more energy-efficient than running them on the CPU. As with most Nordic Semiconductor solutions, current consumption plays an important role in the Axon NPU. It is optimized for battery-powered devices, which makes the nRF54LM20B a perfect fit for:

  • Wearables
  • Health monitoring
  • Security systems
  • IoT applications

The list of features offered by the Axon NPU and leveraged by the nRF Connect SDK is growing. The main benefits and typical usage of the accelerator are:

  • Fast and power-efficient execution on many models with little to no CPU intervention (due to Smart DMA)
  • Axon supports 8 quantized models with an option for int32 model output with configurable radix
  • Up to 2 inputs per node
  • Maximum tensor size is: 1024 (height), 1024 (width), 512 (channels)
  • Native Activation Functions: ReLU, ReLU6, and LeakyReLU
  • CPU activation functions that are not natively supported by Axon: Softmax, Tanh, and Sigmoid
  • Acceleration of DSP vector operations: square root, natural log, exponent, and FFT

Axon accelerates typical operations:

  • Conv1D and Conv2D (special acceleration for pointwise)
  • Depthwise Convolution (without channel multipliers)
  • Fully connected
  • Vector Add
  • Average Pooling (without padding)
  • Max Pooling
  • Mean
  • Strided slice, concatenation, and splitV

The image below shows the architecture of the Axon NPU accelerator.

Edge AI Add-on for nRF Connect SDK

Nordic Semiconductor offers an Edge AI Add-on for nRF Connect SDK, including a range of tools and samples to compile and run the models on the Axon NPU or a PC simulator. The add-on includes:

  • Axon NPU driver – Provides access to the Axon NPU through the Axon driver API
  • Axon Compiler – Compiles TensorFlow Lite models to be used by the driver
  • Test applications – Presents selected Axon NPU functionality
  • Samples – Complete sample applications using Axon NPU
  • Simulator – The software that executes the applications and provides performance estimates without the hardware access needed

Model training and execution

The Axon NPU can execute a model from the following sources:

  • TensorFlow – Models can be executed on the Axon NPU and need to be compiled using the Axon Compiler, which is a part of the Edge AI Add-on for nRF Connect SDK
  • Nordic Edge AI Lab – A cloud-based Nordic AI platform that can be used to design, train, and optimize models. The platform produces models that are compatible with the Axon NPU and can be executed using Edge AI Add-on for nRF Connect SDK

It is also possible to create models that can be executed on the CPU, although this is most suitable for simpler models. It applies to both TensorFlow models and those being produced by Nordic Edge AI Lab. To run a TensorFlow model on the CPU, the littleRT runtime is needed. The most efficient option will be using Nordic Neuton custom models, which are up to 10x faster and more energy efficient than running TensorFlow models.

Edge development workflow

 

   

Switch language?

Progress is tracked separately for each language. Switching will continue from your progress in that language or start fresh if you haven't begun.

Your current progress is saved, and you can switch back anytime.

Register an account
Already have an account? Log in
(All fields are required unless specified optional)

  • 8 or more characters
  • Upper and lower case letters
  • At least one number or special character

Forgot your password?
Enter the email associated with your account, and we will send you a link to reset your password.