The development of artificial intelligence has brought convenience and changes to people’s lives. However, a cloud-centric architecture is not always the ideal solution in every situation, such as concerns about information security and challenges in product design due to power consumption. As a result, embedded artificial intelligence technology, characterized by its distributed nature, has begun to receive attention, and its future development will enable devices to have higher intelligence.
1
Embedded Artificial Intelligence Technology Gains Attention
Currently, most artificial intelligence computations are performed in data centers, i.e., running in the “cloud”. However, with the advancement of technology, people have discovered a huge opportunity emerging away from the data center at the internet edge—embedded artificial intelligence is gaining increasing attention.
The Internet of Things has a vast number of terminal devices. In the future, if the data captured by these network nodes needs to be uploaded to the cloud for intelligent processing or deep learning, it will pose a significant challenge to network bandwidth. “Another challenge is power consumption. Many devices are powered by batteries, such as smart mobile devices and new energy vehicles, which place increasingly high demands on device power consumption,” said Wang Junfeng, deputy director of the Comprehensive Marketing Department of Renesas Electronics (China) Co., Ltd. Therefore, embedded artificial intelligence technology characterized by edge computing has begun to gain attention.
In response, Wang Fan, technical director of ThunderSoft, stated that like cloud computing, the role of edge computing is also to optimize resources and improve efficiency. For example, some small embedded devices perform basic information collection and processing at the edge, meaning that after sensor data is transmitted to the gateway, data filtering and processing occur, eliminating the need to send every piece of raw data to the cloud, saving a significant amount of cost. Cloud-based AI aims to solve problems better, while embedded AI aims to solve problems more economically.
2
Mobile Computing Vendors Accelerate Layout
In March of this year, ARM released the DynamIQ technology for artificial intelligence applications. At a recent technology forum, ARM showcased a new processor based on DynamIQ technology: the Cortex-A75 processor, Cortex-A55 processor, and Mali-G72 graphics processor. Nandan Nayampally, ARM’s vice president and general manager of the Computing Products Division, stated, “We need to empower distributed intelligence that is faster, more efficient, and more secure from network nodes to the cloud.” The Cortex-A series processors using DynamIQ technology can achieve 50 times the artificial intelligence performance compared to devices based on Cortex-A73 after application optimization, and can improve the response speed between the CPU and designated hardware accelerators on the SoC by up to 10 times.
At the recent GMIC (Global Mobile Internet Conference), Meng Pu, chairman of Qualcomm China, also highlighted the topic of artificial intelligence. He stated, “In the future, machine learning will develop in coordination between the cloud and the end. We do not believe that all artificial intelligence is implemented in the cloud due to issues such as personal privacy, information security, and transmission latency. Qualcomm’s flagship processor Snapdragon 835 has high-performance graphics processing capabilities, along with a digital signal processor (DSP) and software algorithms, which will enable the terminal to achieve machine learning capabilities. I believe that the development of artificial intelligence and machine learning at the terminal will accelerate in tandem with that in the cloud.”
In response to the increasing demand for edge computing, NVIDIA launched the new Jetson TX2 development board, which integrates the entire artificial intelligence system onto a single circuit board, allowing Jetson TX2 to better run deep learning functions on terminal devices, thereby developing more intelligent devices. Compared to the previous generation Jetson TX1, the performance of Jetson TX2 has doubled, while its power consumption is less than 7.5 watts, improving energy efficiency by more than twice.
According to reports, the reVISION stack technology launched by Xilinx features reconfigurability and all forms of connectivity, allowing developers to fully utilize stack technology for rapid development and deployment of upgrade solutions, which is crucial for developing intelligent vision systems for future needs. Moreover, this technology enables developers to gain significant advantages when combining machine learning, computer vision, sensor fusion, and connectivity applications. For example, compared to other embedded GPUs and traditional SoCs, reVISION improves machine learning inference performance to 6 times per watt, and computer vision processing speed to 42 times per watt per frame, while latency is only 1/5.
3
Autonomous Driving and Data Security Will Be First to Implement Applications
Currently, embedded AI has begun to enter the market, particularly penetrating and applying rapidly in the fields of autonomous driving and data security.
“If the assisted driving system computes in the cloud, the data collected by the device must be uploaded, computed, and then returned to the terminal, this will inevitably introduce some latency, and in driving scenarios, this latency means an increased risk,” said Zhao Kun, deputy director of the Automotive Electronics Department of the Application Technology Center at Renesas Electronics (China) Co., Ltd. At the same time, data security is also a key point of concern, as uploading to the cloud increases the risk of privacy leakage.
Therefore, embedded artificial intelligence will find extensive applications in edge computing products. According to the recently released “Edge Computing Industry Alliance White Paper,” there are three development stages of edge computing.
The first is connectivity. Achieving massive, heterogeneous, and real-time connections of terminals and devices, automatic network deployment and operation, while ensuring connectivity security, reliability, and maintainability. Remote automatic meter reading is one of the application scenarios that solves the problem of a large number of electric meters.
The second is intelligence. Introducing data analysis and business automation processing capabilities on the edge, intelligently executing local business logic, which can significantly improve efficiency and reduce costs. Predictive maintenance of elevators is one such application.
The third is autonomy. By introducing artificial intelligence, edge computing can autonomously perform business logic analysis and computation, and dynamically and in real-time complete self-optimization and adjustment of execution strategies.
Reprinted from China Electronics News Author: Chen Bingxin
