In response to the explosion of edge AI, traditional embedded manufacturers are on a new journey

Recently, at Computex 2023, relevant leaders in the processor field from TI, NXP and ST introduced their respective companies' understanding of the future of embedded systems, especially the edge AI field, and their respective companies' response plans.

Texas Instruments: Edge AI vision processing empowers the future possibilities of embedded systems

Sameer Wasson, Vice President of the Processor Division of Texas Instruments, gave a speech on "Edge AI Vision Processing Empowers the Future Possibility of Embedded Systems". He said that a comprehensive embedded processing product portfolio should have three major elements: Higher integrated perception capabilities ; Popularize more AI and make it easier to use in embedded systems.

Wasson said that the development of embedded systems requires a balance between cost and development difficulty, as well as collaborative optimization of software and hardware to achieve the best design results. In addition, embedded system developers prefer portable and reusable software and hardware designs, so a platform strategy is crucial.

TI has three major advantages in the field of edge AI, including a highly integrated and scalable edge AI processor portfolio, easy import of artificial intelligence and machine learning functions into existing applications, and open source tools and software stacks to assist AI development, without even requiring engineers to develop them themselves. Any coding can add AI functionality to the system.

This year, TI launched six embedded vision processors based on Arm Cortex, including AM62A, AM68A and AM69A processors. The computing power ranges from 1TOPS to 32TOPS and supports from one to up to 12 cameras.

Since TI launched the AM335x and widely introduced the 64-bit processing concept into industrial applications, Arm has begun to enter a wider range of industrial fields.

In AM6x, TI strives to lead the industry from price to power consumption, from development threshold to scalability.

NXP: Edge AI needs more security

Ali Osman Ors, global director of AI and ML strategy and technology at NXP, emphasized the security considerations of edge AI.

According to an IBM report, manufacturing is the most attacked industry globally in 2021, with ransomware still being the culprit, accounting for 23% of attacks. In the future, as smart factories continue to evolve, security issues will emerge even more.

Ali emphasized that machine learning requires all-round defense, which includes not only code and equipment, but also a lot of critical data. He cited several protection methods, including defending against adversarial attacks, preventing data poisoning, preventing model theft, performance monitoring and model protection.

IP is an important part of machine learning. Regarding the intellectual property rights of machine learning models, if the classification is based on factual elements such as "cat/dog", "car/pedestrian/traffic light", etc., it is difficult to judge whether the training data set can be claimed. Copyright, as this does not include any creativity. However, in the industrial or medical industries, such as developing a unique set of image diagnostic models, some unique encryption methods are needed to prevent theft.

NXP has introduced the eIQ Model Watermark tool into the eIQ toolkit for machine learning development, adding watermarks to the machine learning method. Developers can select specific types of images with secret graphics to combine to generate trigger images, and the Watermark tool can expand the original training data based on the trigger images. The user chooses to label the triggering image with a "watermark category" that is distinct from the actual category of the underlying image, e.g. labeling a triggering image that is actually a cat as "dog". Training with this extended training set produces a model with unique features on trigger images, called "Mountweazels". This is the watermark of the machine learning model. When an independently trained model uses a trigger image, the resulting classification is the actual category of the underlying image of the trigger image, but the original trained machine learning model and systems that copy the watermarked machine model will be classified into the "watermark category". This suggests that the model plagiarized the original model.

And the NXP eIQ model watermarking tool has been optimized to not affect model performance or accuracy.

Regarding products, NXP has launched a number of new products in the i.MX9 series this year, using the Cortex A55 core, and including independent MCU-like real-time domain, Energy Flex architecture, advanced security and Dedicated multi-sensory data processing engine (graphics, image, display, audio and speech).

EdgeLock is a preconfigured security subsystem that simplifies the implementation of complex security encryption technologies and helps designers avoid costly mistakes.

Facing the future, Ali believes that generative AI and quantum computing will bring unprecedented impact to cryptography. To this end, NXP is continuing to innovate. For example, in 2022, the National Institute of Standards and Technology (NIST) selected the Crystals-Kyber professional algorithm co-signed by NXP for the formulation of post-quantum cryptography standards.

STMicroelectronics: Edge AI can bring higher energy efficiency

Arnaud Julienne, vice president of STMicroelectronics Asia Pacific Microcontroller and Digital IC Product Division, (MDG) Internet of Things/Artificial Intelligence Technology Innovation Center and Digital Marketing, emphasized the role of edge AI in energy saving and consumption reduction.

Julian said that electricity consumption in residential and commercial buildings can account for 90% of that in large cities, with major electricity consumption including lighting, HVAC, home appliances and other applications. STMicroelectronics is improving power waste through the digital technology revolution in various fields. For example, it helps improve the energy efficiency of washing machines from D-level to A-level, uses BLDC to replace AC motors, improves HVAC efficiency by 30%, reduces TV standby power consumption and supports LED lighting, etc.

Julian gave an example of a weighing application on a washing machine. Using the STM32G4 MCU equipped with the edge AI algorithm and the SLLIMM IPM chip to measure the current during the rolling and rotating process, the clothes can be accurately weighed without sensors. Compared with traditional weighing In the heavy mode, the accuracy is increased by three times, which can make the motor run more accurately and save more electricity and water resources. This algorithm, which STMicroelectronics calls Zero Speed ​​Full Torque, also ensures that the current is smaller when the motor is started, thereby further saving power.

Another example is the use of edge AI for arc detection in the photovoltaic power generation process. Using the AI ​​function of STM32, the detection accuracy can be improved by 99% compared with traditional arc detection.

In 2019, STMicroelectronics released STM32 cube AI, which has now become the most popular AI development tool in the embedded field. In 2021, STMicroelectronics released NanoEdge AI, which has a large number of built-in AI library functions including the above-mentioned arc detection, weight detection, etc., allowing engineers without any AI skills or even data to develop AI products. In 2023, STMicroelectronics released the Cube AI cloud service to further simplify the development process.

This year, STMicroelectronics released the first MCU with NPU, STM32N6. Its neural network acceleration (ST YoloLC NN) capability is 75 times higher than that of STM32H7, and it has image functions such as MIPI, ISP and H.264, as well as STSafe security elements. .

In terms of MPU, STMicroelectronics has released the second-generation industrial 4.0 edge AI microprocessor STM32MP25, which uses the Arm Cortex-A35 core and supports TSN.

Julian also emphasized STMicroelectronics’ portfolio in wireless connectivity. In addition to Bluetooth, Sub-1GHz and UWB, STMicroelectronics has also developed ST60, a high-bandwidth, low-power innovative wireless device based on 60GHz millimeter wave technology. Connectivity technology.

Finally, Julian said that as the demand for MCUs becomes increasingly strong, STMicroelectronics is investing extensively in internal production capacity and actively expanding partners to ensure the supply of MCU production capacity in the future.