Technical Articles
2025-01-15

Empowering Intelligent Control with Lightweight Embedded AI

Artificial Intelligence (AI) is not a new technology concept; its history can be traced back to the 1950s. Entering the 21st century, AI has made significant breakthroughs in fields such as image recognition, voice recognition, natural language processing, robotics, and autonomous driving, benefiting from the rapid increase in computing power and the fast development of technologies like big data, cloud computing, and deep learning. Especially in recent years, with the continuous expansion and penetration of applications, concepts such as “Edge AI” / “Embedded AI” / “On-device AI” have gradually become industry hotspots.
Although the names differ, these various concepts essentially describe the same thing – “embedding AI algorithms on the device side, enabling devices to possess intelligence, automation, and efficiency capabilities.” This trend is most evident in the AI solutions of leading microcontroller (MCU) manufacturers worldwide.
Comprehensively Embedding AI on the Device Side
2023 is a pivotal year for the development of artificial intelligence. In the technology sector, the development of large models and generative AI has further triggered changes in computing paradigms, industrial momentum, and the landscape of computing power services; in the application sector, enterprises are accelerating their transition from business digitization to business intelligence; in the chip sector, the variety of chips is gradually enriching from server GPUs to edge MPUs to lightweight end-side MCUs. This means that as the demand for artificial intelligence rapidly grows, a comprehensive AI architecture characterized by “cloud-edge-end” has already formed.
Currently, developers of end-side products and applications are increasingly adopting AI and machine learning (AI/ML) algorithms to match and recognize complex patterns, helping to analyze data and make decisions accordingly. As the most end-side digital logic device directly controlling material actions, MCUs should also be AI-enabled and should not be excluded from AI. In fact, the use of AI/ML technology is growing extremely rapidly. In terms of numbers, from millions to tens of millions of AI servers, it is expected that by 2030, the shipment of end-side embedded AI devices will reach 2.5 billion, indicating huge market potential.
By adding these intelligent features and bringing them as close as possible to the key operational locations, end-side AI can reduce device latency and improve the accuracy of real-time operations, making human-machine interaction, status monitoring, and predictive maintenance in smart factories smarter and more energy-efficient; it can be used to optimize the battery management systems of new energy vehicles and adjust vehicle states to adapt to drivers’ habits, ensuring vehicles are driven in an energy-efficient manner; and it can leverage networks composed of millions of smart sensors and IoT nodes to improve monitoring, resource management, assist citizens in smart cities, and enhance logistics with autonomous drones and vehicles.
Cross-Industry Integration is Challenging
Many network edge applications that can utilize AI/ML capabilities need to operate under extremely stringent power consumption constraints, with many devices needing to run for extended periods, even months or years, on just a single charge or solely relying on collected and stored energy. On the other hand, in the past decade, AI models have developed rapidly, and the number of innovations in AI classification models has been increasing year by year, leading to the emergence of new implementation methods, which require optimization in hardware and algorithms.
At the same time, due to the performance limitations of MCU hardware, the complexity of AI software, high real-time requirements of industry applications, strict energy consumption limitations, and high data security requirements, the industry application of embedded AI on MCUs requires business teams to have rich AI knowledge and experience, as well as embedded software and hardware capabilities.
On one hand, the development team must engage in deep cooperation with the implementation scenarios. Only by using the foundational data based on real scenarios and the customized solutions provided by the embedded AI team can we truly solve problems and create value.
On the other hand, it is essential to achieve rapid deployment of AI models to MCUs. General AI models are often not designed for embedded applications, and the variety of AI development tools is vast. However, the RAM, Flash resources, and CPU computing power of MCUs are usually very limited. Choosing the right AI model and development tools to develop models that solve specific scenario problems, and deploying AI models to resource-constrained MCUs, is a complex engineering challenge.
A²MCU Embedded AI Empowers Various Industries
As a chip company committed to enabling intelligent terminals for interconnected devices and serving as the foundation for digitalization, connectivity, intelligence, and low carbonization, HiSilicon is deeply integrating ultra-lightweight technology frameworks, extreme performance inference requirements, and convenient rapid deployment capabilities into MCUs through the A²MCU solution, providing new choices for MCU industry customers to explore intelligent applications.
The name A² signifies “two As multiplied, creating an exponential cumulative effect.” One A represents “Application Specific,” embodying HiSilicon’s customer-centric approach that closely integrates chip design with customer application scenarios; the other A represents the application of “lightweight embedded AI technology” in MCUs and embedded fields.However, it should be emphasized that unlike MCU solutions with built-in NPUs, HiSilicon’s solution focuses more on achieving AI training and inference integration on limited computing power and configuration of RISC-V CPUs.
Empowering Intelligent Control with Lightweight Embedded AI
HiSilicon’s entire series of MCUs provides an ultra-lightweight AI technology framework capable of meeting AI training and inference requirements across multiple scenarios. It can quickly convert multiple models into code and import them into projects for developers to facilitate rapid product deployment.
1) Minimal Framework: AI models deployed on MCUs, after being converted into network layer running code, directly call optimized operator libraries of the RISC-V core, eliminating the need for model parsers and other generally complex frameworks. The open-source architecture of RISC-V supports custom instruction sets, enabling better support for the optimization implementation of operator libraries, which is a key advantage of RISC-V over other cores.
Empowering Intelligent Control with Lightweight Embedded AI
2) Extreme Performance: This means simplifying the training and inference processes while ensuring scenario benefits. This includes but is not limited to: optimizing training models, including model structure optimization, reducing memory read/write and computation; quantizing the model after training to make it smaller and inference faster. Lightweight operators, memory optimization, and deep performance tuning: balancing the reduction of computation and memory access overhead through operator library lightening, pre-rearranging operator data, and reusing memory, reducing multiplication operations, and minimizing memory access costs.
Empowering Intelligent Control with Lightweight Embedded AI
Lightweight model pruning and compression
Empowering Intelligent Control with Lightweight Embedded AI
Quantizing after training, making models smaller and inference faster
3) Easy Development and Deployment: HiSilicon’s A²MCU embedded AI solution offers various model conversion options, such as models developed through TensorFlow Lite, PyTorch, MindSpore, etc., which can be quickly and conveniently converted into code and imported into project code.
Empowering Intelligent Control with Lightweight Embedded AI
In addition, HiSilicon’s A² solutions have also achieved innovations in OS and algorithm integration:
4) Deep OS Integration: Recognizing the value brought by the deep integration of chips and operating systems, HiSilicon’s A² solutions deeply coordinate and optimize MCU chips and OS that emphasize high real-time performance with those requiring high-performance real-time computing. Generally, to ensure high real-time performance, existing solutions in the industry often do not use operating systems. However, this lack of basic scheduling functionality leads to a significant increase in code complexity and subsequent maintenance difficulty once the MCU code exceeds ten thousand lines. HiSilicon, through joint innovation with openEuler, has developed a hybrid deployment solution that can run on limited MCU resources, called UniProton+BareMetal (no OS bare metal). This solution requires minimal hardware resources, running with just 4KB RAM and 4KB Flash. By deploying this hybrid solution, the priority and real-time performance of existing high-real-time tasks are maintained, and it can run directly in a BareMetal environment; at the same time, for tasks with lower real-time requirements, multi-threaded task management can be achieved through a scheduler, providing multi-threaded management capabilities, reducing the complexity of code development for developers, and facilitating easier post-maintenance and modifications for customers, as well as cross-chip porting.
5) Industry Customization: From the current application landscape, as the degree of industry application segmentation continues to rise, the advantages of industry-specific MCUs are evident. For example, in the fields of motor control and power control, specialized MCUs dominate to meet the energy products’ pursuit of higher energy efficiency and higher power density. Coupled with the rapid growth of advantageous industries such as China’s home appliance industry, industrial sector, energy, and new energy vehicles, local chip companies are growing rapidly. Statistics show that in 2022, the domestic MCU market was approximately 40 billion yuan, with domestic MCU sales accounting for about one-fourth of this market, indicating significant potential in sectors such as consumer goods, home appliances, industry, and energy. HiSilicon’s A² solutions emphasize algorithm optimization and customization for scenarios, such as the “A²MCU, making air conditioning more energy-efficient” solution, which is a preliminary showcase of A²MCU’s capabilities. This solution learns complex working conditions based on air conditioning environments, operations, and target parameters, improving overall energy efficiency during the operational cycle through embedded AI algorithms. The deep integration of business scenarios and AI reinforcement learning models brings differentiated competitive advantages in energy saving for air conditioning products, ultimately achieving a 16% reduction in energy consumption during the temperature adjustment phase.
Empowering Intelligent Control with Lightweight Embedded AI
Conclusion
Overall, transferring AI data processing from the cloud to edge deployment requires a series of innovative semiconductor technologies, including ultra-low power consumption technologies and system methods. The combination of these technologies reduces system power consumption and bandwidth requirements while further enhancing the computational efficiency of the new generation of microcontrollers designed for edge devices. However, for developers, this means that they must find the “critical balance point” among multiple factors such as performance, efficiency, safety, and total system cost from the very beginning of MCU design. Without sufficient technical accumulation and rich application experience, achieving innovation and creating incremental value in low-power AI scenarios is certainly not an easy task.

Related posts

Leave a Comment

Your email address will not be published. Required fields are marked *