The nRF54LM20B System-on-Chip with large memory capacity integrates an NPU that accelerates TensorFlow Lite-class models up to 15 times faster and with significantly lower power consumption than running on an Arm Cortex CPU. Furthermore, it delivers up to 7 times the performance and 8 times better energy efficiency than the closest AI edge solution on the market, making AI workloads with high-frequency sensors, audio, and events viable on devices powered by small batteries.
“This new generation of AI-enabled capabilities at the edge radically transforms what small, battery-powered devices can perceive and interpret in the real world,” says Øyvind Strøm, EVP Short-Range at Nordic Semiconductor. “By bringing powerful intelligence directly to the device, without compromising energy efficiency, latency, or system complexity, we are enabling a new class of products that are much smarter, faster-responsive, and significantly more energy-efficient than any previous solution.”.
By simplifying how AI models at the edge are created, optimized, and deployed, Nordic makes it easier for developers to add meaningful intelligence to their products without increasing energy consumption or development complexity.
Nordic is strengthening its edge AI offering by combining advanced hardware acceleration, flexible model support, and intuitive development workflows to make on-device intelligence viable in consumer products.
The new nRF54LM20B System-on-Chip with NPU delivers the performance needed for real-time intelligence in battery-powered wearables, smart home devices, smart audio, industrial and medical sensors, and tracking devices.
This enables:
accurate activity detection,
natural response to sound and presence,
instant anomaly identification, and
motion interpretation—
all without the need for cloud processing.
Simplifying AI Development at the Edge
In addition to new hardware acceleration, Nordic has expanded its ultra-low-power AI solution with features that remove the usual barriers to integrating intelligence into connected, battery-powered IoT products:
Developers can build AI models for direct on-device execution by loading custom datasets and defining the model architecture to generate classification or regression models. This simplifies the design of AI architectures from scratch for wearables, industrial sensors, and smart home devices.
Developers can generate a custom wake word model from a single typed phrase (e.g., “Hello Nordic”), without having to collect and label thousands of audio samples. This dramatically reduces the cost of creating voice datasets and enables voice controls to be implemented in smart speakers, wearables, remote controls, and battery-powered audio accessories.
Models created with the new online tool Nordic Edge AI Lab can be optimized for both accelerated inference on the NPU and efficient execution on the CPU, offering flexibility based on performance and power requirements.
This facilitates integration between models, tools, and embedded firmware, while enabling applications such as
audio classification,
anomaly detection,
advanced IMU processing,
predictive maintenance,
environmental sensing,
and signal processing.
