Chips become the key to AI development. From the cloud to the edge is the trend

Benefiting from the continuous improvement of semiconductor technology and the continuous optimization of algorithms, the development of AI-related industries can be described as leaps and bounds. From the perspective of AI chips, according to the MarketsandMarkets report, the AI chip market will continue to expand at an annual growth rate of about 40%; while faster instruction cycles, lower power consumption, and deep learning algorithms are the most important factors at this stage. main requirement.

Cloud computing: Oligible market share, major players are major players

In the cloud computing market, CPU (central processing unit) and GPU (graphics processing unit) are mainly used due to the need to process huge data in data centers and supercomputers. NVIDIA is the earliest in the layout of AI chips, and its GPU has long been the dominant player, but it has not stopped. The familiar Ampere is still selling well, and the new generation of Hopper is ready to go on sale in the third quarter of this year; , NVIDIA is also building the first high-performance CPU-Grace using LPDDR5 (LwoPower Double Data Rate, low-power memory), which is expected to be available next year. Although the CPU giant Intel (Intel) entered AI chips late, it has completed the integration of CPU, GPU, FPGA (programmable logic gates) by successively acquiring Nervana System, Habana Labs, Movidius, Mobileye, Altera, eASIC, etc. Array chip) and ASIC (special application chip) comprehensive layout, its ambition in the AI chip market should not be underestimated.

On the other hand, network giants and IT manufacturers that originally belonged to downstream buyers have further developed their own chips because they are heavy users and hope that the chip functions better meet their own needs. For example, the TPU (Tensor Processing Unit) launched by Google is the It belongs to the representative of ASIC; while Microsoft cooperates with Intel on AI chips based on FPGA architecture, so that the Azure cloud computing platform can better reduce the computing delay. It is worth mentioning that Graphcore, a British AI chip unicorn founded in 2016, whose innovatively developed IPU (Intelligence Processing Unit) combines “learning” (training algorithms through a large amount of data) and “inference” (executing algorithms). , interpreting terminal data) functions on the chip at the same time, not only has received capital injections from world-class companies such as Microsoft, Samsung, Dell, etc., but is also regarded as a strong enemy of NVIDIA GPU.

Edge computing: the next wave of development, everyone has a chance

In the past, AI "inference" work was mostly carried out in the cloud, but based on the market's criticism of its high cost and delay, as well as end users' requirements for privacy, in recent years, AI "inference" work has gradually been devolved to the edge, such as micro Data center, terminal equipment, etc. Since the edge AI chip processes data between the cloud and the terminal device, it focuses on the real-time, low power consumption, and small size of on-site data interpretation and transmission. Therefore, FPGA and ASIC (such as TPU, NPU, VPU, BPU, etc. are used) ) type-based.

The evolution from the cloud to the edge means facing thousands of industries and a small number of diverse and scattered application scenarios, which has also contributed to the strong market demand for edge AI chips. In 2024, the shipment of edge AI chips may exceed 1.5 billion, with a compound annual growth rate of at least 20%; Tractica also predicts that the market size of edge AI chips will be 3.5 times higher than the cloud AI chip market by 2025. On the other hand, the rise of edge computing has also created huge business opportunities for market players (such as many AI chip startups).

As far as the security monitoring industry is concerned, the phenomenon of AI shifting from the cloud to the edge (terminal equipment) has gradually become apparent in recent years. Taking the network camera with AI intelligent recognition technology (hereinafter referred to as AI Camera) as an example, because many emergencies in the monitoring environment often exceeded the range that the AI camera database and algorithm can handle, even if the back-end image database is expanded, it is still impossible. Comprehensive and lack of real-time. When the edge AI chip is built in the camera, the ability of intelligent recognition and self-learning is greatly improved, which not only reduces the misjudgment rate, but also exceeds the ability of human eyes to recognize, and can respond to various emergencies in a more real-time and comprehensive manner. .

From technical competition to complete solution, AI intelligent recognition is gradually "landing"

In addition to the better use of AI intelligent recognition technology in terminal equipment such as surveillance cameras and DVRs, the solutions for different application scenarios in various industries have also made great progress. From the "secutech 2022 Taipei International Security Technology Application Expo", it can be observed that AI intelligent recognition technology has almost become the "standard" of security control manufacturers, but it has not been focused on AI algorithms, models, recognition targets, recognition rates... and other technologies Instead, various vertical application solutions are highlighted and emphasized in the most eye-catching positions of various exhibition booths and posters and billboards. Compared with the previous situation where most security monitoring manufacturers are still forming teams and designing solutions, or only a few solutions have been implemented, in the past two years, due to cross-border integration and continuous construction and expansion of the industrial ecosystem, manufacturers have accumulated There are more and more practical application cases, more and more aspects, and more and more confidence.

Combined with access control management and traffic applications for epidemic prevention, the management plan for large-scale construction sites/contractors has increased

As the related variant viruses derived from COVID-19 continue to ravage the world, even though most countries have given up the "clearing" thinking and are forced to adopt the practice of "coexisting with the virus", epidemic prevention is still an indispensable measure in daily life. Therefore, the access control (entry and exit) management system that combines face, mask recognition and body temperature detection functions is still the most common AI application solution in the market, and can be seen almost everywhere; especially the one that can be quickly and easily combined with the employee attendance management system, it is even more subject to government administration. Institutions, commercial buildings, industrial plants are welcome to the vast number of enterprises. And the integration of AI humanoid, object, behavior and other recognition functions, safety management solutions for temporary workers and outsourcers in construction sites/factories have also increased significantly at this year's venue; such solutions can not only quickly identify jobs with AI intelligent recognition technology Whether the personnel are fully worn in accordance with the regulations (such as masks, helmets, reflective vests, safety shoes, etc.), and can also define dangerous behaviors and dangerous area police boundaries (electronic fences) according to the needs of the owner. Once the safety system is violated, an alarm will be triggered; If combined with license plate recognition, construction site security requirements, etc., a more complete solution can be formed.

On the other hand, the penetration rate of license plate recognition systems in parking lots has become higher and higher. In addition to the entrance and exit gates, the AI cameras equipped with indoor parking spaces can not only display lights, but also have license plate recognition and vehicle detection functions. Built-in black/white list. It can also be seen at the exhibition that the AI edge computing function has been fully demonstrated in the application of road traffic - the AI camera integrated with AI recognition technologies such as license plates, objects, and behaviors can not only immediately recognize license plates, colors, and types of vehicles, but also It can calculate traffic flow, track objects and detect scattered objects; for the existing cameras that have been set up at a large number of intersections, there is no shortage of corresponding solutions - just add AI computing methods to the original intersection equipment, and the control center can adjust the traffic lights according to the traffic flow. seconds, keep track of illegal/parked vehicles, and even provide real-time data on self-driving cars. In the safety and management applications of rail transportation, there are also AI cameras integrated with vehicle DVRs, and AI patrol solutions with touch-integrated interfaces. Train drivers can see the situation of the entire train at a glance as long as they are in the cab, saving the need for a The hassle of cruising around with boxes. In addition, the identification of marine ships at the port, the identification of drones, and the intrusion of humans and animals (such as birds) at the airport, etc., also have corresponding plans on display.

Diversified applications bloom

In the residential area, I believe everyone is familiar with terms such as "cloud smart house" and "smart community". With the blessing of AI recognition technology, combined with face recognition, body temperature detection, access control and intercom systems, license plates are used. The identified parking lot management system, as well as the energy and environment automatic detection system, can be further integrated with the property management system to provide security personnel with a more intelligent way of personnel scheduling and management, and it can also make it easier for on-site security personnel to use the App. , Complete daily work in real time. In the application of retail stores, AI's intelligent analysis of people flow, customer base, and hot spots is familiar to everyone. This year's emphasis is on how specific people (such as lost children or the elderly) can use cross-camera tracking and search functions. Quickly lock its path; this function can also be applied to personnel search and route analysis in various public areas.

In addition, it is emphasized that AI limb detection without facial features (considering the right to privacy) can be applied to rehabilitation, sports, etc., to better judge whether the relevant actions are in place? Fall detection can also be used in medical/long-term care institutions to observe whether a patient/elderly falls, allowing medical staff to respond in real time. The detection of smoke and flames by AI intelligent recognition has also been extended to glasses or smart billboards with AR (Augmented Reality) function, which is more conducive to the judgment of disaster rescue scenes.

Conclusion

In the AIoT era, no AI chip architecture can be applied to various scenarios. This is an opportunity for the development of edge AI chips. Similarly, an intelligent recognition model trained by an AI platform may not be applicable to various scenarios. Opportunities for security monitoring manufacturers to deepen the AI vertical market. Edge applications are diverse, and we also expect that various security-related solutions combined with AI intelligent identification technology can, with the joint efforts of security monitoring manufacturers and related ecosystems, be more refined and closer to user needs. Application fields are blooming everywhere.