AXONS — Neural processing unit

nRF54LM20A | nRF54LM20B Datasheet

Axon NPU is an ultra-low power neural network accelerator built for artificial intelligence (AI) and machine learning (ML) applications. It enables real-time processing while keeping power consumption to a minimum, making it perfect for IoT edge computing where battery life is critical.

Note: Axon NPU is only available in selected device variants. For more information on Axon NPU device availability, see Product overview.
The main features of Axon NPU are the following:
  • Fast and power efficient execution of neural network operators
  • End-to-end execution with little to no CPU intervention for most operators
  • Support for quantized int8 models.
  • Flexible DMA and execution pipeline that can support many functions beyond traditional neural network operators
  • Compatible with a software tool chain and driver that stay up to date with growing Axon NPU functionality
  • Acceleration of int24 DSP vector operations square root, natural log, exponent, and FFT
Figure 1. Block diagram
Block diagram

Optimized for battery-powered devices, Axon NPU is a high-performance yet energy-efficient solution for wearables, security systems, health monitoring, and other IoT applications. By reducing power usage, it extends system battery life without compromising processing speed, ensuring reliable performance even in resource-limited environments.

Axon NPU is intended for use with the libraries in Nordic Semiconductor device SDKs. Base operations include convolution (1D and 2D), Fully Connected, Add, Multiply, and Pooling (average and max). See the Axon NPU toolchain documentation for a complete list of limits and currently supported functionality.