Nordic’s Integrated SoC to Simplify Edge AI for Battery-Powered IoT

Nordic Semiconductor recently introduced its nRF54L series to bring hardware-accelerated edge AI to battery-powered industrial and household devices. The nRF54LM20B uses a wireless SoC with a large memory footprint and an integrated Axon NPU, enabling it to run AI workloads directly on the device. With tools like Nordic Edge AI Lab and compact Neuton models, developers can create production-ready AI that runs locally, keeping power and memory use to a minimum.

Nordic’s nRF54LM20B integrates an Axon NPU and low-power wireless SoC to enable on-device AI in battery-powered systems. Image used courtesy of Nordic Semiconductor

The platform is built around three key needs for engineers working with constrained devices. First is the Axon NPU, which provides dedicated hardware acceleration, so AI inference can run without overloading the main MCU. Second, the nRF54L series delivers up to 7x the performance and 8x the energy efficiency compared to previous solutions, translating into longer battery life and faster response times. Finally, Nordic Edge AI Lab makes it easier for developers to generate ultra-small Neuton models that fit within the tight memory limits of MCUs.

Axon NPU and Hardware Acceleration

The Axon NPU is a built-in accelerator in the nRF54LM20B that handles AI inference directly in hardware, running quantized neural operations in parallel and using local memory and direct data movement to efficiently move inputs and outputs.

Block diagram of the nRF54L. 

By pushing heavy AI processing onto Axon, the main Cortex-M33 processor is freed up to focus on control logic, wireless communication, and I/O tasks rather than managing inference. The NPU also includes 2 MB of nonvolatile memory and 512 KB of RAM, providing enough space for larger AI models and working data to help keep the system responsive by reducing competition for memory and processing resources during operation.

A 7× performance increase means the device can finish its AI work much faster, so it spends less time in higher power modes. An 8× increase in energy efficiency reduces the amount of battery used each time the AI runs, making it realistic to keep sensing tasks running more often, even on small batteries. By processing data locally instead of sending it to the cloud, the system also uses less radio power and responds more quickly, which is important for applications that need fast responses or must keep data private.

Neuton Models, Edge AI Lab, and System Integration

Neuton models are very small AI models, often under 5 KB, that can run directly on a microcontroller. They’re designed for tasks such as spotting unusual behavior, recognizing simple gestures or activities, and basic biometric monitoring. Nordic Edge AI Lab handles much of the heavy lifting by preparing data, building and optimizing models, and exporting them to run on either the CPU or the Axon accelerator. This reduces the need for manual tuning and makes it easier for embedded teams to add on-device AI without deep machine learning experience.

The nRF54LM20B integrates multiple system components into a single device, including Nordic’s Axon platform, a 128-MHz Arm Cortex M33, a RISC-V coprocessor, high-speed USB, and up to 66 GPIOs. It also includes an ultra-low-power, 2.4-GHz radio supporting Bluetooth LE, Matter, Thread, Zigbee, and proprietary protocols.

Performance data and low sleep current figures indicate the device's low average power consumption in typical battery-powered applications. Nordic also links the on-device AI capability to its broader cloud services, enabling these models to be updated over the air, monitored in the field, and managed across device fleets, which supports long-term tuning, compliance, and secure updates without hardware changes.

Ultra-Efficient AI Acceleration

Support is available through Nordic Edge AI Lab, with Neuton models already running on nRF54 series devices. The nRF54LM20B is currently in sampling with selected customers, but wider availability is expected in early Q2 2026. The platform is intended for applications such as audio and event detection, anomaly monitoring, and gesture recognition, where local processing, low latency, and battery life are key considerations.