Nordic Semiconductor is set to buy U.S. startup Atlazo, primarily for the company’s AI hardware IP and expertise. The IoT SoC maker based in Oslo, Norway, plans to add in-house hardware AI acceleration IP to devices across its portfolio, Kjetil Holstad, executive VP of strategy and product management, told EE Times.
“While Nordic customers today are using [standalone AI accelerators] alongside Nordic SoCs, Atlazo represented a chance to get hold of technology that we can start to embed in our devices,” he said.
The purchase still needs regulatory approval. But if all goes well, Atlazo IP could begin appearing in Nordic products in as few as 12-18 months, Holstad said.
Nordic’s product range includes wireless microcontrollers (MCUs) with ultra-low power Bluetooth, ANT, Thread and ZigBee connectivity for applications like wearable medical and fitness devices, but the company has also expanded into cellular and WiFi chips in recent years.
“I don’t see a limitation on where [Atlazo’s AI accelerator] technology can fit,” Holstad said. “It might not be the same implementation in all products, but I do think it belongs across our technologies and across what Nordic does. If we are thinking about IoT, this technology is coming, full steam ahead.”
Atlazo, founded in 2016 in San Diego, Calif., has eight team members who will join Nordic.
Nordic last year announced the acquisition of leading-edge static SRAM startup Mobile Semiconductor, whose low-power, low-leakage memory IP was already in all Nordic’s wireless IoT SoCs. Nordic also previously acquired Imagination Technologies’ WiFi IP and team in 2021.
“Anyone can go to TSMC and use their memories, but then everyone who is doing that performs the same,” Holstad said. “Memories, acceleration and connectivity—all of these are crucial for getting the ultra-low power aspects right. For us, it’s about figuring out how we can be better than everyone else and that’s how both [Mobile and Atlazo] acquisitions came in.”
Atlazo’s first-generation hardware IP for AI acceleration, Axon 1, can accelerate many types of neural network, including those used for the medical and wellness sensor types, which are critical to Nordic’s customer base. Axon 1 can also handle neural networks for vision, voice and vibration sensing, including LSTM (long short term memory) and GRNN (general regression neural networks), as well as feature extraction techniques like MFCC (mel-frequency cepstral coefficients).
Axon 1 is silicon-proven in the startup’s AZ-N1 SoC, which launched in 2021. Nordic will not continue to sell the AZ-N1, but will integrate next-generation Axon IP—already under development—into its products.
From Nordic’s perspective, it’s important to have this IP in-house to drive its development, Holstad said.
“What’s core for Nordic is the ultra-low-power aspect, that’s where we strive and where we differentiate, but also on the ease of use,” he said. “We could have licensed from or collaborated with others, but we think this is going to be essential to customers and we want to own that technology so we can really nail the ultra-low power aspects of it.”
Holstad added that Nordic’s customers are solving more and bigger problems with Nordic chips than ever before.
“We went through a journey of becoming an IoT company, first and foremost,” he said. “Now we’re starting to look at expanding within IoT and within ultra-low power, extending the footprint of problems we can solve on behalf of our customers…. We’re now doing multi-core devices that have a lot of computational power, more like an MCU with radio than a radio with an MCU.”
With the field of tinyML gaining traction across the industry, AI algorithms are becoming small and lightweight enough to run on general-purpose Arm MCU cores. Existing Nordic customer applications for tinyML typically include anomaly-detection and sound-detection applications.
Nordic’s existing AI software stack is based on Arm’s ML libraries, and the company partners with third parties, including Edge Impulse, for the best possible developer experience, Holstad said.
The Atlazo purchase comes with some sensor-specific neural network algorithms, which may be offered as Nordic’s in-house models in the future.
Overall, however, the company also wants to allow its customers to differentiate, depending on their level of experience and resources, he added.
“If IoT is going where we think it’s going, it needs more [AI compute],” he said. “You cannot just have a stream of all the sensor data in the world up to the cloud and let AWS solve things; that’s not sustainable, for so many reasons. Our customers need to solve problems as close to the sensor as possible, and we now see a need to go beyond using classic general-purpose Arm [compute] to solve that problem.”
Adding hardware acceleration for IP also makes sense for applications that fit into the Arm core today, Holstad said.
“If an inference takes 100 milliseconds on an Arm core, it might take [a fraction of that] on an accelerator, so you’ll save a lot of power,” he said. “So if the cost-value proposition is right, it will be included [across all Nordic product lines]. There’s always the exception to the rule, but right now I think we can make a good value proposition for including [AI acceleration] in pretty much everything we are thinking about doing in the future.”
Atlazo also has sensor interface and power management IP alongside general expertise in ultra-low power design, which would be added to Nordic’s toolbox for “solving the world’s battery problems,” Holstad said.
If regulators approve the purchase, Atlazo’s San Diego, Calif., office would become Nordic’s third U.S. outpost, together with the Mobile Semiconductor office in Seattle and Nordic’s existing base in Portland, Oregon.
Source : EE Times
