Edge AI on Nordic
On-device machine learning with Nordic Semiconductor's Axon NPU, the nRF54LM20B SoC, and the Edge AI Add-on for nRF Connect SDK.
Nordic Semiconductor brought hardware-accelerated machine learning to the ultra-low-power BLE world in early 2026 with the nRF54LM20B — the first Nordic SoC to integrate a dedicated Axon Neural Processing Unit (NPU). This section is your guide to building production firmware that runs on-device inference (keyword spotting, audio classification, sensor analytics, anomaly detection) without round-tripping to the cloud.
The Axon NPU originated from Nordic's 2023 acquisition of Atlazo (San Diego) — a startup specialising in always-on AI processors and energy management for tiny edge devices. The architecture has since been integrated into Nordic's own silicon and is being rolled out across the wireless portfolio (next stop: the nRF92 cellular-IoT modules).
What you'll find here
nRF54LM20B SoC
The first Nordic SoC with an integrated NPU — full specs, memory layout, radio capabilities, and how it differs from the NPU-less nRF54LM20A sibling.
Axon NPU architecture
What the Axon NPU is, the operations it accelerates natively, performance numbers (GOPS, energy efficiency), and quantization expectations.
Development tools
Edge AI Add-on v2.0 for nRF Connect SDK, the Axon Compiler, Nordic Edge AI Lab, Edge Impulse integration, and Neuton custom models.
Getting started
End-to-end workflow: pick a model, compile it for Axon, integrate with your Zephyr application, and flash to the nRF54LM20 DK.
Why on-device inference matters for nRF designs
- Latency — Wake-word detection, gesture recognition, and keyword spotting need millisecond-class response times that a round trip to a cloud model cannot meet.
- Power — Keeping the radio off most of the time and waking on a local inference event is dramatically cheaper than streaming sensor data.
- Privacy — Audio and biometric data stay on the device.
- Connectivity independence — The product still works when Wi-Fi or cellular is unavailable.
The Axon NPU lets you do all of this inside the same coin-cell-class power budget that defines the nRF54L Series — something traditional Cortex-M-only inference workloads typically can't sustain.
How FirmwareMaestro fits in
FirmwareMaestro generates Zephyr-native scaffolds that already wire up the nRF Connect SDK toolchain. For Edge AI projects, that means:
- A
prj.confthat pulls in the Edge AI Add-on (nrf_edgeai_lib) and the Axon NPU drivers - A Devicetree overlay for the on-board microphone / IMU / sensor source
- A
main.cskeleton that initialises the NPU, loads a compiled model header, and runs an inference loop in a dedicated Zephyr work queue - Generated PRD, architecture, and HAL documents that explicitly call out the inference budget, model footprint, and DSP pre-processing path
Start with the Getting Started workflow.
Power Management & Optimization
Techniques and strategies for minimizing power consumption in battery-powered embedded devices, from sleep modes to radio duty cycling.
nRF54LM20B SoC
The first Nordic SoC with an integrated Axon NPU — 128 MHz Arm Cortex-M33, 2 MB NVM, 512 KB RAM, BLE 6.0 / Matter / Thread / Zigbee, and high-speed USB.