Nordic Semi buys TinyML assets to boost embedded AI roadmap
Cette publication existe aussi en Français
Nordic Semiconductor has bought the assets of US TinyML tool developer Neuton.ai for an undisclosed sum to boost its embedded AI roadmap, which also includes wireless chips with a hardware accelerator later this year.
Neuton.ai has developed a Web-based toolchain that takes in training data to provide a compact machine learning model, typically under 5Kbytes that can run on an embedded microcontroller core, hence TinyML. The deal includes 13 engineers in Pleasanton, California, and the IP.
“This is a technology acquisition for us. We bought the assets, the IP and the people. We see this as applicable across all the end markets we address,” Nordic CEO Vegard Wollan tells eeNews Europe.
“Many of our customers are researching solutions in this space but only a few have moved to manufacturing and we believe this offering will be a low threshold way to enable edge AI in a low deployment way which we find particularly appealing.”
Nordic sees the tool being used with its latest nRF54 series of Bluetooth wireless chips that are based on the ARM Cortex-M33, although the inference models will run on any of the ARM cores in Nordic devices for health trackers, sensor fusion, smart home and industrial applications.
“There is a lot of talk around AI accelerators but this means you don’t need the hardware accelerator as customers can create models in a web-based interface to create an inference model and the code,” says Oyvind Strom, EVP short-range at Nordic.
“This is done through patented algorithms with neuron to neuron algorithms rather than convolutional neural networks. This technology can do any kind of sensor fusion, gesture and keyword recognition but not voice recognition or vision. There are a lot of application targets in smart home and wearables that you can use with software ML.”
“We are highlighting the nRF54 as this is the latest platform but the Neuton products can run on all our MCUs in the nRF52 and 53 as well as in our long range and WiFi devices which can all benefit from day one.”
“The threshold to using AI tools has been high in the past and the beauty of the Neuton acquisition means you can start with a limited data set, upload more data to improve without having to develop large models in TensorFlow. There are customers using the Neuton models training then and exploring a path to products, but not out on the market yet.”
“We are in production with the 54 but only a limited number of customers,” said Wollan. “We are expecting that to increase in the second half of this year as the traction with the 54 series is extremely strong.”
Hardware AI accelerator
Back in 2023, Nordic also acquired Atlazo, a team of 8 people in the US developing a hardware AI accelerator. That AI accelerator will be part of a chip launched later this year, says Wollan.
“With the Atlazo IP we are well covered with the accelerators which are competing with the ARM accelerators but more targeted for our end applications. We are very comfortable with what we have covering our application base so we don’t see a need for the higher end accelerators right now. Typically our design cycles are 18 to 24 months so the first products will come later this year,” he said.
The Neuton acquisition also opens up an interesting option for improving the performance of existing devices with embedded AI added through over the air (OTA) updates.
“All our products can be updated over the air, there is no technical restriction. Its an interesting thought and if our customers would like to do that it is very feasible,” said Strom.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
