Silicon Labs' Product Manager covers Silicon's AI SoCs

Let’s start with an introduction of you and your position, as well as a brief introduction of Silicon Labs

My name is Matt Maupin and I have been working on low power wireless technologies for over 20 years, even before the IoT was a thing. I have been involved in Product Marketing and Product Management for 2.4 GHz technologies including Wi-Fi, Bluetooth, 802.15.4 and even sub-GHz technologies. I came to Silicon Labs in 2012 as a Product Manager soon after the acquisition their acquisition of Ember.

Back then, I saw Silicon Labs’ commitment and vision to low power wireless solutions and the IoT and it has been great being part of the execution of that vision.

Today, Silicon Labs is a leader in secure, intelligent wireless technology for a more connected world. Our integrated hardware and software platform, intuitive development tools, unmatched ecosystem, and robust support make us an ideal long-term partner in building advanced industrial, commercial, home, and life applications. We make it easy for developers to solve complex wireless challenges throughout the product lifecycle and get to market quickly with innovative solutions that transform industries, grow economies, and improve lives

Before we talk more about the technology’s applications, could you provide an initial introduction of Silicon Labs’ MG24 and BG24 families of 2.4GHz wireless system on chips (SoCs) as well as its latest software toolkit?

The BG24 and MG24 families are Silicon Labs’ two new families of Matter-ready, ultra-low power 2.4 GHz wireless SoCs. They feature the industry’s first integrated AI/ML accelerators, incorporate PSA Level 3 Secure Vault protection, and have the largest memory and flash capacity in the Silicon Labs portfolio. The BG24 supports Bluetooth Low Energy, Bluetooth Mesh, and proprietary, while the MG24 adds support for support for Matter, Zigbee, OpenThread, and multi-protocol operation.

These SoCs feature the highest on-chip TX power, RX sensitivity, large memory, and extensive GPIO, making them ideal for esh applications. MG24 is the optimal SoC for Matter-enabled high-performance smart home devices. BG24 brings high accuracy, low current consumption, nd the highest PSA Level 3 security to medical and industrial sensor applications.

Silicon Labs’ new software toolkit is designed to allow developers to quickly build and deploy AI and machine learning algorithms using some of the most popular tool suites like TensorFlow. The toolkit ensures that developers have an end-to-end toolchain that simplifies the development of machine learning models optimized for embedded deployments of wireless applications.

The solution works with the existing Series 1 and Series 2 wireless SoCs as well as the newly introduced EFR32BG24 and EFR32MG24 products, which feature integrated AI/ML hardware acceleration.

Could you tell us about the artificial intelligence (AI)/machine learning (ML) accelerator technology in Silicon’s BG24 and MG24 SoC families?

The built-in AI/ML hardware accelerator enables up to 4x faster processing of machine learning algorithms with up to 6x lower power consumption while offloading the main CPU for other applications – enabling smarter and faster IoT edge devices with long battery life. On-device AI/ML algorithm processing is required for advanced IoT applications; this minimises latency and traffic over the internet for time-sensitive applications  and strengthens privacy and security since data and processing stay at the edge.

Using Silicon Labs’ solution, designers can take advantage of built-in ML hardware acceleration and lower power consumption to achieve greater accuracy and speed. It’s exciting to see what’s possible without sending data to the cloud.

Off the back of that last question, could you describe the ways in which the SoCs support Zigbee, OpenThread, Bluetooth Low Energy, Bluetooth mesh, and more?

The MG24 and BG24 SoCs are a multiprotocol devices supporting mesh protocols including Bluetooth mesh and 802.15.4 protocols like Zigbee and Thread.  RF performance is often overlooked for mesh technologies.   For example, when covering large mesh installations, one may think I only need to get to the next device and so on to expand coverage.

While this is true, having better RF performance has advantages in a mesh network. First, you cannot guarantee what coverage looks like, so you could have gaps in the mesh. The other issue, is if you have poor performing RF devices, it will generate more hops in the network.

This creates more traffic and results in longer latency. That is why we have focused on providing the best RF performance for our devices. For example, for 802.15.4 which includes Zigbee and Matter using OpenThread, we have a link budget of 124 dB, this reduces any potential for gaps in the mesh means less traffic and retries in the network.

Devices from other companies would require a separate PA and/or LNA to get this performance, adding cost and complexity to the design. The other key aspect of this device is the amount of Flash and RAM required to run some of these more complex mesh networking. For example, OpenThread and Matter will require more then 1024K of Flash and 128K of RAM. With 1536K of Flash and 256K of RAM, this device has the headroom to support these protocols today and tomorrow.

As the number of IoT devices increase, the potential of AI and ML is stronger than ever – but product designers who deploy such technology at the edge face the challenges in terms of performance and energy demands. Could you describe how the BG24 and MG24 SoCs overcome such challenges?

IoT product designers see the tremendous potential of AI and machine learning to bring even greater intelligence to edge for applications like home security systems, wearable medical monitors, sensors monitoring commercial facilities, industrial equipment, and more. But today, those considering deploying AI or machine learning at the edge are faced with steep penalties in performance and energy use that may outweigh the benefits.

The BG24 and MG24 alleviate those penalties as the first ultra-low powered devices with dedicated AI/ML accelerators built in. This specialized hardware is designed to handle complex calculations quickly and efficiently, with internal testing showing up to a 4x improvement in performance along with up to a 6x improvement in energy efficiency vs using the Cortex for these calculations.

Because the ML calculations are happening on the local device rather than in the cloud, network latency is eliminated for faster decision-making and actions. In addition, RF transmit and receive current typically have the highest impact on battery life. Allowing the decision-making to be local reduces the transmit and receive activity, thus improving battery life.

Would you like to describe how the SoCs benefit smart homes, security, and other IoT-related areas – as well as any other applications that apply?

While we are seeing applications across the board, the majority of the customers are implementing the new SoCs in products for Zigbee, Zigbee/BLE mulitprotocol and of course OpenThread and Matter. Security providers like the low power aspects of the device as well as the wide range of peripherals, but also are looking at how to benefit from the AI/ML hardware accelerator for areas like glass break detection, and so on.

Commercial building companies are exploring how to make their building systems, including lighting and HVAC, more intelligent to lower owners' costs and reduce their environmental footprint while consumer and smart home solution providers are working to make it easier to connect various devices and expand the functionality by using features of the BG24 and MG24 including the 20-bit ADC and AI/ML accelerator, thus improving their user’s experience and making their product more valuable to the consumer.

Let’s now talk about Silicon’s latest AI/ML software toolkit. How is this toolkit designed to help developers in the development of machine learning models that are optimised for the embedded deployments of wireless applications?

The AI/ML software toolkit will allow developers to create applications that draw information from various connected devices – all communicating with each other using Matter – and then making intelligent machine learning-driven decisions.

One application example of Silicon’s AI/ML edge technology that stood out for me was its ability to utilise the combination of data from both movement detection and audio sensors to inform when office lights should be on or off. Could you describe this process with attention to the sensors and data processing at the edge involved?

In a commercial office building, many lights are controlled by motion detectors that monitor occupancy to determine if the lights should be on or off. However, when typing at a desk with motion limited to hands and fingers, workers may be left in the dark when motion sensors alone cannot recognise their presence.

By connecting audio sensors with motion detectors through the Matter application layer, the additional audio data, such as the sound of typing, can be run through a machine-learning algorithm to allow the lighting system to make a more informed decision about whether the lights should be on or off.

I know that just some of the uses include anomaly detection, predictive maintenance, audio pattern recognition for improved glass-break detection, simple-command word recognition, and presence detection/people counting. Could you briefly describe your thoughts on such processes as well as any other applications that you think the machine learning software could apply to?

At a high level, everything is related to pattern recognition. As a device is operating, it generates data (whether that's audio data, visual data, or data around other operating parameters like temperature), and over time that data generates a pattern that shows what normal operation looks like, and that pattern is what trains the particular machine learning model.

What ML devices at the edge are doing, is running the data they're collecting in real time against that trained model and looking for anomalies in the pattern, such as the sound of breaking glass, or specific patterns like a wake word, as defined by the trained model. So while the applications you mention are what we are seeing today, there are really limitless potential applications this could apply to as the number of data-collecting IoT devices continue to expand.

Following on from that last question, I know that over 40 companies have already taken an interest in developing and testing Silicon Labs’ BG24 and MG24 SoCs. With particular attention to the applications and the technology involved, could you describe what you believe has most attracted these customers to the technology?

We have over 50 companies now evaluating and designing their products based on the BG24 and MG24 SoCs.  These companies have been drawn to the BG24 and MG24 platforms by their ultra-low power, advanced features like AI/ML capabilities and the highest level of IoT security with our Secure Vault technology. Another key factor is the large Flash and RAM help to future proof the device as protocols like Matter evolve over time.

Our customers recognise this and see the MG24 and BG24 SoCs as something they can build around regardless of their operating environment.

I appreciate that the BG24 and MG24 SoCs are considered by Silicon Labs to be its most capable family of systems on chips to date. With reference to previous SoCs for comparison, could you describe what makes the xG24 systems stand out for you?

The BG24 and MG24 are part of our 2^nd generation of EFR32 wireless SoC devices (EFR32xG1, xG12, xG13 and xG14).  While we are still having tremendous success with the first generation devices because of key features that are still in Series 2 like excellent RF performance including up to +20 dBm output power.

However, Series 2 devices has made some improvements across the board. For example, such as an ARM Cortex-M33 processor capable of up to 78MHz. At the same time, we have reduced active current consumption by up to 50% for the MCU as well as receive and transmit.  Secure Vault technology used in our Series 2 provides the most advanced security features for a device in this class, protecting from both remote and local cyber attacks. In addition to Series 2 improvements, the BG24 and MG24 have several unique benefits.

This device has the largest Flash and RAM combination in our portfolio, making it ideal for complex mesh protocols and application including multiprotocol and even Matter. This is the only Wireless SoC for edge devices that has an integrated 20-bit ADC, allowing our customer to eliminate the need for an external high accuracy ADC, or even better, they can pull in functionality that they have not done previously due to the cost of an external high accuracy ADC.

We have already seen this with some of our Alpha customers for this device.  And as we have talked about previously, our new is our AI/ML accelerator on this device that reduces both the computation time and current for machine learning inferencing at the edge of the network. We have shown up to 4x faster inference performance and 6x lower power consumption vs Cortex-M. This is going to become increasing more important as we see more demand for running AI/ML at the edge versus the cloud.