News broke recently of Nordic Semiconductor acquiring Neuton.AI in what appears to be its bid to more deeply integrate Edge AI into its devices. To understand the decision-making behind this acquisition and Nordic’s more general approach to AI, IoT Insider spoke to Kjetil Holstad, EVP of Corporate Strategy at Nordic.
The company is known for its low-power Bluetooth devices but it also has a portfolio of microcontrollers (MCUs), a market where devices have evolved over time – with greater capabilities, more computational power, and more functionality required all driving this change.
At the same time, around five years ago, Nordic’s customers began to experiment with machine learning and neural networks to improve the intelligence of their own products. Due to Edge AI’s ability to process data on the ‘edge’ of the device with greater bandwidth and lower latency, it is extremely appealing for applications like wearables, where processing data in the Cloud is too power-hungry and not sustainable for a battery-powered device.
“It started with understanding that pure MCUs are not enough,” said Holstad. “Our customers were starting to put neural networks and machine learning capabilities onto our MCUs.”
Taking this into account, in 2023 Nordic acquired Atlazo, an AI and ML company, who specialised in hardware accelerators. The acquisition of Neuton.AI brought software capabilities on top of that.
There are two interesting aspects to Neuton that Nordic noticed: firstly, it has a patented neural network framework, 10 times more efficient than comparable open-source models like TensorFlow, subsequently reducing the code size and the power consumption, and secondly, it has a web-based platform that allows you to upload your data, pre-process it, extract features, and creates a small neural network that can be run on the device again.
Close alignment
Developers can face constraints on running ML models on the Edge of the device: namely, the size of the device prohibits it, and building custom ML models is complex. Nordic hopes to address this through acquiring Neuton and its capabilities.
“Neuton, we got to know earlier, because they are in this constrained, embedded side of [applications],” explained Holstad.
In other words, Nordic recognised in Neuton its understanding of running ML models on the Edge on small, constrained devices. This is important because Nordic’s chips are used in smaller, more compact devices, not just in wearables, but in wireless headphones and keyboards too.
It’s important to differentiate what Nordic is doing in AI compared to what other companies are doing in this space, who might be looking at large language models (LLMs) and generative AI, both of which consume more power and require larger devices, generally speaking.
Nordic has chosen to focus on TinyML, which applies machine learning models to resource-constrained devices, notably MCUs, with optimised models to run efficiently without using too much memory or power.
“TinyML and TensorFlow and others have taken it to a certain level,” said Holstad. “So they could start to fit on Nordic-type devices, but perhaps they spent 50% of the available computational power on our chips. What Neuton is doing is taking that a step further, where they reduce the size tenfold.
“So now we’re going from a megabyte, perhaps to 500 kilobytes, but with Neuton, the average framework size is … five kilobytes … so we can run on all [of] Nordic’s portfolio from day one.”
Neuton’s patented technology can automatically generate ML models under 5KB in size, which can be deployed across 8-, 16-, and 32-bit MCUs.
This was a conscious decision from the company as part of a broader effort to democratise ML on the Edge, to ensure it could be run across a wide range of devices, and to avoid its customers from being locked into using certain hardware.
The announcement mentioned combining Neuton’s models with Nordic’s nRF54 series of SoCs because it is new and that is where the company sees a lot of design activity, but the plan is to run it on legacy, existing and future devices of Nordic, Holstad said.
Going forward
Embedding AI and ML in electronics is still very much in its infancy, and as a result, Nordic do receive a fair few questions from customers.
“A lot of the conversations we have is around telling people what they can use this technology for,” said Holstad. “Neuton … have a lot of examples [of] applications where they show what types of capabilities they can do, whether that’s for human machine interfaces or hand washing tracking or parcel tracking.”
Having these conversations ultimately provides the opportunity to educate people on how AI can help them.
“The aim here is to make sure that people can build better products,” Holstad said aptly. “What we’re hoping to do here is for anything that is Edge AI, machine learning at the Edge, we take care of all the complexities of training models, selecting models, whether that should be running on the MCU or an accelerator.
“Taking all that complexity away, and building it into our strong, easy-to-use brand,” Holstad concluded.
There’s plenty of other editorial on our sister site, Electronic Specifier! Or you can always join in the conversation by visiting our LinkedIn page.