Interview with ADI： exploring the application boundary of AI MCU and integrating CNN hardware accelerator will be the technical trend of edge AI processing.

Author | Wei Shiwei

**

Thanks to the drastic changes in market supply and demand in recent years, as well as the new growth of Internet of Things and automotive electronics market, MCU (Micro Control Unit) has gradually become a hot potato.

MCU, also known as the "brain" of electronic system, can not only control other parts of the system according to certain procedures, but also make processing, calculation and decision by collecting external or internal data, which can be widely used in the fields of consumption, industry, medical care and automobiles.

To put it simply, whether it’s the lifting control of the car window, the temperature adjustment of the air conditioner, or the hot blood oxygen, heart rate and even blood pressure measurement products, the realization of its functions is inseparable from MCU. It has already penetrated into all aspects of people’s lives and is closely related to people’s food, clothing, housing and transportation.

According to IC Insights data, in 2021, due to the tight supply chain, the average selling price of MCU increased by 10%, and the sales scale reached a record $19.6 billion. It is estimated that the global MCU sales will increase by 10% to $21.5 billion in 2022, among which the growth of the automobile MCU market will exceed that of most terminal markets.

In the analog chip market, although ADI started with its high-precision signal chain chip products and developed a powerful business empire together with its power supply products, it also has considerable technical strength in the MCU field. Since 1995, ADI has shipped more than 1 billion pieces of MCU products. meanwhileFrom 2020, AID began to develop the marginal AI MCU product line on the basis of traditional MCU.It can help battery-powered devices to realize artificial intelligence and Internet of Things applications more easily.

On the whole, ADI MCU products have the characteristics of low power consumption, rich interfaces, new communication modules, complete evaluation schemes and development examples, simple product development and high security. Depending on the functional application,ADI’s MCU products can be divided into three categories:

• Low power MCU:It has the characteristics of small volume, low power consumption and large storage, and can be applied to industries, Internet of Things, medical care, wearables and other fields.
• Secure MCU:It has a secure system architecture and strong anti-attack encryption ability, and can be applied to smart phones or terminals with high security requirements, such as POS machines and card readers.
• Artificial intelligence MCU:Born out of the first kind of low-power MCU, it is characterized by its ability to push AI reasoning from the cloud to the edge, which can be applied to smart home, face punching, voice control and other applications.

Overview of ADI microcontroller products

Around MCU products, ADI also provides a series of support resources to simplify the user’s design, including professional support team, easily integrated driver routines, generous and concise mechanical design, robust and reliable evaluation suite and clear and easy-to-read manual guide, which greatly simplifies the customer’s R&D process.

At the same time, in terms of hardware, ADI has also released a variety of evaluation kits, development versions and reference designs, as well as a wealth of software libraries and design documents, compilation and debugging IDE, etc. It also supports specialized platforms such as KEIL, IAR, mbed and eclipse, and has set up an online and offline technical support FAE/ADI China Technical Support Center, which can respond to customer needs in time.

Talking about the application of MCU in the marginal field, Xin Yi, a senior engineer of ADI’s China Technical Support Center, believes that because the interconnection between things has generated a huge demand for data processing, only powerful computing power can be competent, so IoT technology is deeply integrated with AI, which has spawned the concept of AIoT, in which IoT is equivalent to the neural network of people everywhere, and AI is equivalent to the human brain. To achieve intelligence similar to human beings, The device must perform a lot of matrix operations, which puts higher demands on the storage space, computing power, data interaction speed and cost of the device, and only large servers deployed in the cloud can be competent.

However, the interconnection between IoT devices needs battery power supply, and the data flow between devices cannot completely rely on the cloud, so IoT applications also need low power consumption and low cost, which also brings challenges to the integration of AI technology and IoT technology.

In view of these pain points, ADI combines the advantages of both AI and IoT to realize AI reasoning task at the edge of IoT devices, so that devices can make their own calculations and decisions locally, without having to connect to the Internet. Compared with cloud AI,Edge AI has the characteristics of good real-time, low bandwidth requirements and high privacy, and it also has the same AI common characteristics as cloud AI.

Technical advantages of ADI edge AI microcontroller

For example, ADI’s edge AI solution MAX7800X series consists of two microcontroller cores (ARM Cortex M4F and RISC-V) and a convolutional neural network (CNN) accelerator. The architecture is highly optimized for the edge, and the microcontroller core is responsible for data loading and starting, while the AI reasoning is specifically responsible for the convolutional neural network accelerator. Based on the division of labor and cooperation of the two hardware, the MAX7800X series does not need networking, but also supports battery power supply, which greatly meets the requirements of edge AI.

In addition, in the aspect of low-power MCU, ADI integrates the functions that traditional multi-MCU have through a single chip, and has built-in power management module, which has the advantages of ultra-low power consumption, high performance and abundant resources. At the same time, some products also have built-in Bluetooth module, which can greatly reduce the system size.

In order to further understand the layout of ADI’s MCU product series, 36Kr and other media recently had an in-depth exchange with Li Yong, senior business manager of ADI’s MCU product line. While sharing the details of ADI’s MCU products and business progress, they also shared their views on the intelligence and market trends of MCU from the industrial level.

Q: The oximeter is very popular recently. What does ADI think of this market?

Li Yong: The popularity of oximeter may be due to the COVID-19 epidemic, but in fact, before the occurrence of COVID-19, ADI had a special blood oxygen testing scheme and a chip for testing blood oxygen saturation, and there was also an algorithm specially developed for this test. The exclusive algorithm can be provided to customers by downloading it into the ADI chip. Customers can build their own products by directly using ADI chips and adding ADI’s ADC and algorithms. Therefore, for this application field, ADI actually laid out very early, and we are very optimistic about the medical and health field.

Q: What are the advantages of integrating CNN hardware accelerator? Will this be a technical trend of edge AI processing in the future?

Li Yong: At present, this is a technical trend. Why does ADI integrate CNN? Frankly speaking, it’s because of power consumption. Because many customers originally used general-purpose processors, such as Cortex-M7, which is relatively fast, with a main frequency of 200MHz, but consumes a lot of power when running, if we want to use Cortex-M7 to calculate CNN algorithm, as mentioned earlier, CNN’s algorithm is multiplication and addition of many matrices, the operation time will be very long.

MCU needs to run fully for a long time, and the power consumption is relatively high. However, if it is oriented to a monitoring device, it can be calculated quickly by integrating CNN now, and then it can go to sleep. In fact, CNN has greatly reduced its power consumption. We believe that this is a very important feature in future marginal applications, especially in some special applications, which is why the microprocessor should be a CNN.

Q: Do you need to consider the combination or matching relationship between the neural network accelerator and MCU? Can the performance of CNN accelerator be expanded in the face of scenarios with different computing power requirements?

Li Yong: ADI’s AI MCU is not only CNN with integrated hardware, but also two microcontrollers, one is Cortex-M4F of Arm and the other is RISC-V. The cooperation between them, M4F is mainly for some applications and communication, which has been allocated, and it is also equipped with FLASH and SRAM. RISC-V is a small kernel with a 32-bit reduced instruction set, which mainly works with CNN. When CNN works, it needs to input some original data, such as pictures, sounds, waveforms, etc. After the data is input through the communication interface or the image camera interface, the RISC-V kernel will transport these data to the storage space, and then let CNN use them. After hardware calculation, CNN obtains some eigenvalues from the original data by matrix multiplication and addition, and then puts them in memory for comparison. It is such a process, so cooperation is definitely needed, but ADI has already planned it, and there are some mature cases that make it very simple for customers to develop.

Q: Max 78000 is a general-purpose MCU with integrated AI functions. What do you think of the relationship between it and AI SoC chips specially designed for a certain type of application or highly adapted to a certain application scenario?

Li Yong: ADI’s MAX78000 is also a relatively small SoC, because it not only integrates the processor core, but also integrates memory (including FLASH and SRAM), so there is no need for external expansion. The memory used in the whole application is integrated, as well as the power module, communication module and some camera interfaces. So it can run some Free RTOS, but it can’t run Android and Linux, but it can run some relatively simple RTOS, so it is also a SoC.

As for the difference, I think MAX78000 is more suitable for edge monitoring and control. It can quickly analyze the original data and perform some control. In the traditional sense, AI chips are relatively powerful, large in size, and may have strong computing power, but they are also very expensive. They are aimed at more applications with high computing speed and strong computing power. This traditional GPU may be used on the server side, while ADI faces the edge side.

Therefore, if you want to use a large FPGA or GPU on the edge, one is that the cost is unbearable, and the other is that there is no way to use batteries for power supply, and it is unnecessary. For example, a camera security camera often only takes a picture and analyzes it once, and then takes a picture and analyzes it again, so it doesn’t need high computing power. Therefore, MAX78000 is very suitable for edge AI applications.

Q: What do you think of the competition from emerging edge voice and image AI chips to ADI MCU? Why doesn’t ADI directly use edge AI chips to deepen these markets, but adopt the concept of AI MCU?

Li Yong: Edge AI emphasizes low power consumption, size, price, and security. Computing power is only one aspect. But for example, to make a camera, one CPU is often enough. This CPU should have both control function and CNN function. This is the current market demand, and ADI can meet it. If you use some traditional AI chips, you may need to add a lot of things outside, including PMIC (power management IC) and some memory, storage and MCU. After all of them are added, the cost is relatively high and the power consumption is relatively high. Therefore, ADI MCU will be a very suitable choice for the edge intelligent market.

Q: By what means are the advantages of low power consumption of many products of Q：ADI realized? Is there room to continue to reduce low power consumption in the future?

Li Yong: I think the wearable low power consumption should be the lowest low power consumption. How to achieve it?

First of all, the Cortex-M4F or RISC-V used in each product itself is low-power, and each chip has different working modes and power consumption modes. For example, in active mode, the chip needs to run at full speed, and the kernel, clock, memory, GPIO and integrator all need to work; In sleep mode, the kernel doesn’t work and stays there, but my clock, memory, GPIO or integrator is still working; In deep sleep mode, the main clock is no longer working, and the kernel is turned off. Some of the memories are charged, because some data need to be kept, while others need to be kept. Some GPIO are basically turned off, and the rest may be activated, and these aspects are combined to reduce power consumption.

Secondly, ADI’s design has multiple clock sources. Our chip has not only 100MHz, but also 7MHz and 4MHz. After our test, the power consumption of low-frequency clock sources has been significantly reduced. In fact, the power consumption of a chip is divided into dynamic power consumption and static power consumption. The dynamic power consumption is that the CPU runs, and the frequency of running is its dynamic power consumption. When it is turned off, there is still a leakage current, which is static power consumption. When using low frequency clock source, the static power consumption can also be reduced.

The third is to improve the integration. ADI integrates power management, communication interface, relatively large memory and even electric frequency conversion chip into the chip. For example, there are two kinds of sensors, one is 1.8V, and the other is 3.3V. Generally, MCU needs to add a conversion chip, but ADI doesn’t, because it has been integrated in the chip and can be configured by software. One port is 1.8V, and the other port is 3.3V. Therefore, after integration, customers can turn off the configuration when it is not needed, so high integration will also help to achieve low power consumption. We put these together to enable ADI’s low-power MCU to achieve wearable low-power consumption.