Silicon Labs Integrates AI/ML Acceleration in Low-power Wireless SoCs
Silicon Labs says that with these two new families, battery-powered edge devices won't face trade-offs with power, protocol support, ML capability, and security.
Today, Silicon Labs released two new families of wireless SoCs aimed at low-power IoT edge processing. BG24 and MG24 wireless SoCs include an integrated AI/ML accelerator for battery-powered edge devices while supporting a wide span of protocols: Matter, Bluetooth Low Energy, Bluetooth mesh, OpenThread, Zigbee, proprietary, and multi-protocol operation. The company has also released an accompanying software platform to help developers to deploy these AI/ML algorithms for a number of use cases.
All About Circuits had the chance to talk with Ross Sabolcik, the vice president of industrial and commercial IoT products at Silicon Labs, to hear about the new SoCs firsthand.
Compute, Connection, and Sensing
These SoCs are meant to be the company’s flagship IoT offerings, and according to Sabolcik, offer the company's best metrics in every facet of the system. “This has our best technology across the board deployed in one device: compute, connection, and sensing,” he remarks.
Block diagram of the MG24 SoC.
With a focus on these three areas—compute, connectivity, and sensing—Silicon Labs claims the MG24 and BG24 SoCs are the first ultra-low-powered devices in the industry to include dedicated AI/ML accelerators.
During internal testing, Silicon Labs found that its new accelerator allowed the SoC to achieve a 4x reduction in power and 6x increase in speed for AI/ML applications. This performance leap has relevant implications for the device’s power efficiency.
“The quicker you can do your compute, the less the device has to be on, and the more battery you can save while you sleep,” Sabolcik explains. “For these SoCs, the AI/ML core has been a really important part of it.”
Both of the SoCs are built around a 78 MHz Arm Cortex M33, which is supported by 256 kB of RAM and 1526 kB of flash.
An evaluation board for the XG24 SoC.
In terms of connectivity, both of these SoCs support 2.4 GHz wireless communication. The major difference between the MG24 and the BG24 is that the former supports BLE and Bluetooth Mesh while the latter supports Matter, OpenThread, and Zigbee. Aside from this difference, both SoCs feature the highest-performing radio that Silicon Labs has ever produced.
The new radio subsystem features a 20 dBm of transmit power while consuming 155 mA TX current, yielding longer ranges of communication or an improved link budget. Beyond that, the radio includes a receiving sensitivity of -97.5 dBm at 1 Mbps GFSK (Gaussian frequency-shift keying) while requiring 4 mA of RX current.
Sabolcik notes, “It's easy to build those types of high-performance radios if you don't care about power. But this device has been designed to be able to operate off of batteries. So we do it with really low current consumption.”
Finally, from a sensing perspective, Silicon Labs has put a 20-bit ADC on their SoC. Silicon Labs tells us that this high-resolution ADC offers the high precision needed in fields like medical and industrial IoT.
Silicon Labs Keeps a Sharp Focus on Security
Bringing AI/ML to the edge, Silicon Labs envisions these devices finding use in basic smart home applications—like gateways and hubs, smart plugs, door locks, and switches—while also offering more complex utility, like wake-word detection, broken glass detection, and predictive maintenance. The two new families are also purpose-built for portable medical devices like pulse oximeters and blood glucose meters.
Block diagram of the BG24.
Involving IoT devices in health and home calls for a high standard of security—and Silicon Labs has also addressed this important consideration into the SoCs' design. Both chips employ Silicon Labs' Secure Vault technology, which protects devices against both remote and local software attacks and local hardware attacks. Secure Vault accomplishes this with a number of features, including a physically unclonable function (PUF), secure boot with RTSL, secure debug unlock, anti-tamper, DPA countermeasures, and a true random number generator (TRNG), among others.
A Supportive Software Toolkit for AI/ML Deployment
In addition to the two new SoCs, Silicon Labs is also releasing a software toolkit to help designers efficiently build and deploy AI/ML algorithms. This toolkit natively supports TensorFlow and can be used with Silicon Labs' Simplicity Studio platform to create new applications and draw data from numerous connected devices that use the BG24 and MG24 families of SoCs. Using the Matter protocol, these devices can communicate with one another to make ML-based decisions.
Consumers demand that their IoT devices balance a unique combination of performance, low power, and connectivity. While this is a difficult trio to blend, Silicon Labs hopes to accomplish this feat with its two new families of wireless SoCs.
“The big challenge with doing AI/ML at the edge is typically twofold. It's either power or bandwidth. If you have a Wi-Fi link and you've got tens of megabytes per second worth of bandwidth (or hundreds of megabytes), you can push a lot of data to the cloud and run learning algorithms there,” Sabolcik explains. “But if you're on a device that doesn't have a lot of power and doesn't have a lot of bandwidth, the ability to make detections [using the BG24 and MG24] at the device is pretty powerful.”
All images, including the featured images, are used courtesy of Silicon Labs.
