Joinwin Electronics HK Limited518031About Join WinJoinwin Electronics HK Limited FLAT/RM 29, 22/F, YAN’S TOWER, 25-27 WONG CHUK HANG ROAD,ABERDEEN HONGKONG Shenzhen ZhongYiYingTong Technology Limited Room 1610, Block A, Overseas Decorative Building, Zhenhua Road,Huaqiang North Street, Shenzhen, 518031Join-Win will be your one-stop purchasing assistant. Join together, achieve win-win!Joinwin Electronics HK Limited518031About Join WinJoinwin Electronics HK Limited FLAT/RM 29, 22/F, YAN’S TOWER, 25-27 WONG CHUK HANG ROAD,ABERDEEN HONGKONG Shenzhen ZhongYiYingTong Technology Limited Room 1610, Block A, Overseas Decorative Building, Zhenhua Road,Huaqiang North Street, Shenzhen, 518031Join-Win will be your one-stop purchasing assistant. Join together, achieve win-win!Joinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics HK LimitedJoinwin Electronics

Home
News
Potential And Challenges Coexist, And The Future Edge AI May Have These New Changes

Potential And Challenges Coexist, And The Future Edge AI May Have These New Changes

March 05 , 2024 0

While edge AI presents great potential and opportunities, developers are also facing challenges from many aspects, including software and hardware, in their implementation. Therefore, it is important to provide guidance and information to developers, and to consider merging and unifying the user experience of various tools to simplify the design process of AI libraries.

Artificial Can Only Develop The Background

"Ai is critical to building the connected world of the future, a smart world of billions of devices that need greater security, connectivity and intelligence, which we call the cloud-connected Intelligent Edge. Matteo MARAVITA, senior manager of ST's Asia Pacific Artificial Intelligence Technology Innovation Center and smart Phone Technology Innovation Center, pointed out in an interview with the media recently that these devices will cover all aspects of our daily life, from families, companies, factories, to cities, buildings, transportation and other fields. When we consider these devices and the business models behind them, we can clearly see what kind of profound impact edge AI will have for the future.

Building Blocks OF The Future World

Security, connectivity, and autonomy are the three basic elements that Matteo lists as needed for edge AI, which together form the core framework of edge AI and provide a solid foundation for the operation of smart devices.

Take vehicle gauge sensors, for example. ST has partnered with HPE Group to optimize electric vehicle motor operation using edge AI - a joint virtual sensor solution that uses AI algorithms running on the Stellar series of vehicle gauge microcontrollers to obtain motor operation data and then infer and estimate unmeasurable temperatures inside the motor. The solution not only runs artificial intelligence, drives motors, but also offers new possibilities for predictive maintenance of electric vehicles to identify possible abnormal states in shock absorption, mechanical and electrical systems.
In the PC sector, ST has optimized the power monitoring of laptops with HP using six-axis IMU MEMS smart sensors. These sensors are able to collect data on the operation of the machine under different conditions, and engineers from both sides have worked together to develop and train AI models based on device and user actions. This model is able to identify specific use cases, such as avoiding overheating and battery depletion before shutting down the laptop, thereby extending the standby time of the laptop and improving its service life.
In the appliance sector, ST is able to estimate the weight of laundry by running an AI algorithm on the STM32 microcontroller controlled by the washing machine motor. According to calculations, a washing cycle saves 15-40% of energy and water. At the same time, the artificial intelligence algorithm can also detect the collision situation that may occur when the drum is rotating, improving the stability and durability of the washing machine.

Matteo believes that the advantages that edge AI brings to related applications can be divided into three categories: the first category is performance enhancement. It goes without saying that thanks to AI technology, the performance of end products will become better and better; The second category is to improve product resistance or robustness under extreme conditions. That is to say, when we encounter some extreme situations that cause the effect of standard algorithms to be unsatisfactory, AI can help the algorithm improve the detection accuracy; The third category is to add new features and new functions in the terminal product that were not previously available to help customers improve the innovation and attractiveness of the terminal product.

The Challenge OF Edge AI

"While edge AI presents great potential and opportunity, developers also face many challenges along the way. On the hardware side, they need to consider security, power consumption and other factors while ensuring the overall performance of the application. On the software side, they need to create, train, deploy and maintain machine learning models in the device." "Matteo said.

ST's AI research and development program began ten years ago, starting from the basic research of machine learning technology, and gradually developed to specialize in AI optimization for edge devices with limited hardware resources:

In 2018, ST released the first STM32 AI library (Cube AI)

In 2020, ST will expand the layout and support of artificial intelligence to automotive, Linux and other fields.
In 2021, the NanoEdge AI ecosystem will be released.
In 2022, a MEMS sensor ISPU with integrated more advanced hardware accelerators was released, capable of running smaller deep learning networks within the same sensor;
In 2023, the STM32 developer Cloud Model Library was released and the Stellar family of MCU tools was updated.

According to the data, STM32 currently ranks first in the number of submitted projects in the MLPerfTiny5 benchmark, with 73% of submitted projects based on STM32 development.

STM32N6 integrated with self-developed NPU is a representative hardware solution of ST. Its machine learning model, developed in collaboration with partner Lacroix for multi-object recognition tracking AI algorithms, showed that the 18-frames-per-second pedestrian, bicycle, and car detection and tracking algorithms were all executed on the NPU, with less than 5% of the load based on the Cortex kernel. On the security side, the product not only integrates the trust zone internally, but also adds the security function peripherals, such as tamper resistance, with these new features, customers and developers can not only protect the data in the memory, but also protect the machine learning model.

Matteo says software presents more of a challenge to developers than hardware. To this end, "providing powerful development tools and support to make it easier for developers to create, train, and deploy models" became ST's core goal. As ST's comprehensive AI development platform, ST Edge AI integrates all the tools and modules needed for edge AI development, from creating use cases with model libraries, to collecting data sets, to deploying models to specific hardware, covering the entire development process, including embedded hardware engineers. All developers, including software engineers and machine learning engineers.

The ST Edge AI suite supports a wide range of deep learning frameworks, including TensorFlow Lite, Keras, PyTorch and other trained machine learning models. It is compatible with the NVIDIA toolkit and provides connectivity to cloud services such as Amazon Cloud Technologies (AWS) and Microsoft Azure, and is compatible with simulation tools such as MATLAB.

"There is currently no universal approach to edge AI development, but it is important to provide guidance and information to developers in the process of developing edge AI solutions. Therefore, the ST Edge AI suite takes this factor into account and can help users find the best methods and tools based on their knowledge and resources. We also consider merging and unifying the user experience of various tools to simplify the design process of AI libraries. Matteo said that both the ST Edge AI and Nano AI Studio are available for free to STM32 customers, while other ARM Cortex-M microcontrollers are also available for free under special licensing agreements.

Big Models And Edge AI

AI large model is undoubtedly the most popular research field in the world. But in Matteo's view, at least for now, large language models (LLMS) are still unable to run on edge AI devices. "We think that in the near future, there will be models, Transformers or LLMS with similar performance but smaller data volumes on the market that may bring a fresher experience to edge AI." 'he said.

From the current actual use situation, the vast majority of artificial intelligence models, including machine learning-based models (such as decision trees, random forests, SVM) and deep learning models (based on fully connected architectures, CNN convolutional neural networks, etc.)RNN recurrent neural networks can run on ST products, the only limitation comes from the complexity of the model and the number of parameters. Therefore, keeping a close eye on the development of the LLM large model and the LLM basic architecture Transformer will be one of the next important directions for ST.

Matteo also mentioned in-memory computing and memory in one technology as an important trend in the development of edge AI. ST believes that "such products will be a game changer for the market" because their performance will be improved to ten times the current level, while consuming only one-tenth of the power.

Of course, innovation is not just in hardware products or platforms, there are also many innovations in software tools and compilers related to NPU and memory computing that require a lot of fine-tuning and optimization.

He stressed that AI developers need to radically shift their thinking from traditional algorithms based on "if-then-else" to AI algorithms that focus on data sets. In particular, there needs to be a shift in thinking when validating overall solutions, with more focus on the definition of use cases, the creation of data sets, and their validation.

At the same time, when it comes to chip design and manufacturing, designing the end product is more important than the chip design itself, as customers need to embed new processes in data set creation, product testing, and manufacturing, and AI capabilities need to be specifically tested before they can be brought to market.