Silicon Labs launches new AI/ML chips, toolkit for edge

Solving some problems

Offerings like these are aimed at addressing some of the problems enterprises face when bringing AI and ML to the edge, said Andy Thurai, an analyst at Constellation Research.

In general, AI/ML edge operations require a lot of algorithms to be run optimized at edges for inferencing, plus they need to be updated frequently as well.

"In general, AI/ML edge operations require a lot of algorithms to be run optimized at edges for inferencing, plus they need to be updated frequently as well," Thurai said.

One reason they're limited is power. Many edge devices use batteries to provide flexibility so they don't need to be connected to a power source.

"If they are not ultra-low power consumption devices, those batteries need to be constantly changed, which can be very expensive to operate," said Thurai

Silicon Labs said its chips are ultra-low power, but Thurai said it needs to be substantiated against other chips on the market.

Another problem with edge and IoT devices is that spotty internet and wireless connectivity have had limited success in bringing AI and machine learning applications to the edge.

Having wireless connectivity out of the box for sensors with Bluetooth and ZigBee can reduce costly implementation and integration processes, since the sensors do not need to be connected to the cloud or a hybrid network, Thurai said.

Silicon Labs system-on-chips

BG24 and MG24 are designed to execute difficult calculations quickly. According to Silicon Labs, the machine learning calculations don't happen in the cloud, but rather on the local device, which make faster decision-making.

"This is true for anyone doing local edge processing," Thurai said, adding that by processing data at the edge, users can avoid cloud round-trip, latency and other problems.

"While many companies claim a faster decision, the cloud round-trip is generally in millisecond, not in seconds," he continued. "For many applications this is not a showstopper."

The Silicon Labs system-on-chips (SoCs) also have large flash and random access memory capabilities, according to Silicon Labs. The vendor said the chips can evolve for multi-protocol support, Matter and trained machine learning algorithms for large data sets.

The systems also include Platform Security Architecture Level 3-Certified Secure Vault, a chip-based subsystem that provides security in IoT devices such as door locks, medical equipment and sensitive deployments.

Secure Vault includes security features that address IoT threats and protects against both hardware and software attacks.

"Security and privacy all go hand in hand when it comes to smart home," said Bob O'Donnell, an analyst at Techanalysis Research. "Having some support for that is going to be important going forward as well."

A new approach

Adding AI on low-power battery devices is a new approach, O'Donnell said.

"We've been hearing all about AI on the edge and things like that. But to be honest, for the most part that's required big devices, like a smartphone," O'Donnell said. "What these chips open up is the possibility to do similar types of things conceptually, with battery-powered devices in smart-home and other kinds of applications."

Since the SoC chips need software to run, O'Donnell said that the accompanying software toolkit is important because many AI programmers are familiar with TensorFlow.

Other than TensorFlow, Silicon Labs also partnered with AI and machine learning tools vendors including SensiML and Edge Impulse for the software toolkit.

Matter matters

Developers can use the SoCs as well as Silicon Labs' Simplicity Studio with the software toolkit to create applications that communicate with each other using Matter.

Matter enables separate IoT ecosystems to come together. For example, in a smart-home scenario, Matter can help unify device ecosystems from Apple, Google and Amazon.

One challenge that Silicon Labs may need to address is the slow evolution of Matter standards, which often get delayed in committees. That can further complicate already complex wireless connectivity tasks for edge and IoT devices, O'Donnell said.

"They've been doing this a long time, so I think they've got this pretty much nailed," he said. "They are combining capabilities they've had in other chips in other chips altogether into this, but that's the way this world works."

"In theory, it should just work straight out the gate," he added.

More than 40 have begun using the chips and the software kit in a closed Alpha program, Silicon Labs said.

The chips and software tool kit will be available for mass deployment in April, the vendor said. Silicon Labs did not provide pricing information.