AI+ Security: GPU and FPGA first come to the future, the main control integrated ASIC

Smart security space: AI products continue to infiltrate, long-term market space of 100 billion

From "visible" to "seeing clearly" to "understand", the demand for intelligent big data analysis is urgent, and the AI+ security trend is obvious. HD technology is getting more and more advanced, and the image resolution is gradually advanced from D1 to 720P, 1080P to 4K, and video surveillance equipment continues to be upgraded in high definition. According to IHS data, the proportion of HD cameras in China increased from 13% to 59% in 2013-2016, surpassing analog cameras for the first time, realizing the transformation of video surveillance from “visible” to “clear” to meet the basic needs of intelligent . High-definition camera production generates massive amounts of data, and traditional manual viewing methods have not met the growing security needs. At the same time, the security field generates a large amount of unstructured data every year. The intelligent processing after massively unstructured data can greatly improve the tracking efficiency. The introduction of artificial intelligence can meet the basic security requirements of post-tracking to pre-emptive prevention. After the high-definition network upgrade in the security field, there is an urgent need for artificial intelligence technology to process massive amounts of data. These have prompted the camera to begin to evolve to "understand", and the trend of intelligent security is obvious.

1

Smart security products first landed in the government market, with a long-term market space of 100 billion yuan. (1) In the short term: Due to the high unit price of AI products and the application of large data for long distances, we believe that the short-term incremental space mainly depends on the public security , transportation and other departments in the government. Assume that the growth rate of domestic/foreign video surveillance industry is 15%/10% respectively. By 2020, the domestic and international video surveillance market will reach 1,683,123 billion yuan respectively. It is conservatively estimated that if the penetration rate of AI products is increased to 10%, then domestic/ The market space of foreign AI products is 168/14.8 billion yuan. (2) In the long run: With the launch of the chip solution with higher cost performance, it is expected that the main control manufacturers such as Hess will inevitably launch IPC main control products including AI exclusive TPU. The security manufacturers headed by Haikang will also develop and launch suitable for themselves. The AI+ chip terminal solution, the unit price of AI products will gradually return to rationality, and the penetration range of smart products is expected to rapidly penetrate into other fields. If the penetration rate of AI products in the future increases to 35%, the global AI product market space will exceed 100 billion yuan.

Existing ecology: GPU is currently the mainstream of AI+ security solutions, industry giants and Nvidia hand in hand

There are two types of “AI+Security” solutions: the front-end solution is the future trend, and the mid-back-end AI solution is the current mainstream. For smart security, there are currently two solutions for front-end and mid-back. The front-end solution is an AI camera solution, which integrates the AI ​​chip into the camera to realize intelligent video capture; the middle-end and back-end solutions use ordinary cameras to collect video information and transmit them to the middle and back end, and insert boards such as GPUs before data storage. The smart server performs a summary analysis. Because the mid-back solution does not need to replace the camera, can process multiple data at the same time, and the deployment cost is relatively low, the algorithm upgrade and operation and maintenance are convenient, and the short-term mid-end solution popularizes faster. In the long run, Haisi and other camera main control chip manufacturers must integrate the exclusive hardware modules for AI calculation inside the chip. After large-scale application, the implementation cost will be drastically reduced. The front-end (smart camera) solution is expected to become the mainstream of smart security in the future. At present, whether it is a front-end or mid-end solution, Haikang, Dahua and other companies are all implemented by NVIDIA GPU (Jetson TX1 product), and the main back-end AI solution.

Security giants work closely with NVIDIA to deploy GPU-based smart device products. At present, the AI ​​algorithms and related products of Haikang and Dahua's two security giants are based on NVIDIA's GPU implementation. In 2016, Hikvision launched a full range of AI products from front-end to back-end, releasing various AIs based on NVIDIA GPU and deep learning technology “Squat” cameras, “Super Brain” NVR, “Facebook” face analysis server, etc. Series products. Dahua Co., Ltd. established the AI ​​Research Institute in the third quarter of 2016. In March 2017, it unveiled a number of “Smart” series front-end and back-end smart devices.

Hikvision: The first security AI+, leading the development of security intelligence. Hikvision began research and development of intelligent analysis technology in 2006, and in-depth study in 2013. With many years of in-depth study and research and the largest R&D team in the industry with more than 9,000 people, the company has won the first place in many image detection competitions including face recognition, vehicle recognition and text recognition. In 2015, the company took the lead in launching the AI ​​center products "Falcon" and "Blade" intelligent server. In 2016, the company first introduced the concept of "Security AI+" at the AMB, and cooperated with Nvidia and Movidius to launch GPU/VPU and deep learning. Technology's "squat", "super brain", "God catch", "Facebook" series AI front and rear products, and incorporate relevant solutions. The company's AI products have been applied to major projects such as Nanchang “Tianwang Project Phase II” and “Belt and Road” Summit Security, which is expected to lead the development of security intelligence.

Dahua shares: Following the layout of artificial intelligence, AI products accelerated. From the beginning of 2015, Dahua began research on artificial intelligence, and developed intelligent technologies such as face recognition, video structuring, abnormal behavior analysis, and high-density crowd analysis based on deep learning. In 16 years, Dahua submitted the test results to the international authoritative face recognition public test library LFW. Dahua-FaceImage face recognition accuracy rate was 99.78%, maintaining the world's first level. In the same period, the company relied on the software and hardware research and development capabilities accumulated on the chip platforms such as CPU, DSP, GPU and FGPA, and developed front and rear face recognition, bayonet alarm, binocular stereo vision, multi-view panoramic mosaic products. A range of intelligent products. In 17 years, the company and Nvidia jointly released a variety of "smart" series of front-end and back-end smart devices and integrated relevant security city project solutions, greatly improving the utilization of video data and promoting the acceleration of big data in the security field.

Future trends: GPU-based cost is a pain point, and future integration into IPC main chip is the trend

Currently, GPU-based smart security is more expensive. At present, the mainstream intelligent security solutions are mostly based on the NVIDIA Jetson TX1 GPU chip. The cost of a single chip is estimated to be around 70~150 US dollars, and the module cost is 200~300 US dollars. Depending on the configuration, each chip can support 2 to 4 channels of video streams, and the cost per channel is higher. Related research shows that Hikvision or the exclusive monopoly of the NVIDIA TX1 GPU chip supply. Other security vendors can only purchase NVIDIA's TX1 modules (that is, GPU boards including chips and storage), and the procurement cost will be higher. GPU-based smart security solutions are expensive and currently hinder the rapid penetration of smart security.

FPGAs are a possible solution to reduce costs by current smart security. In the field of intelligent security, the domestic company Shenjian Technology has cooperated with security vendors such as Dahua and Dongfang Netpower to launch a DPU product based on Xilinx FPGA, which can achieve an order of magnitude improvement in energy efficiency compared to the GPU. %. The solution can control the AI ​​single-channel cost to less than $20, which is cheaper than the GPU solution. At the same time, another domestic company, Horizon Robot, also tried to apply FPGA solutions to security and automotive. Before the advent of dedicated ASIC products, FPGAs are expected to be an effective means of reducing costs in some security scenarios.

ASIC will become the mainstream of security chips in the future. As mentioned earlier, the ASIC-based AI chip solution has the advantages of low cost, low power consumption, and high computing power. Due to the large-scale cost marginal diminishing effect of the chip, after the mass production of the dedicated chip, the cost per chip and related storage cost brought by the AI ​​module is expected to be less than US$2, and the cost of the AI ​​camera using the ASIC solution will be greatly reduced. In other words, it may be possible to convert the camera into an AI camera in the future when the cost increase is very small. We expect that, similar to the AI-specific module integrated with mobile phone chips, Huawei Hi-Ship and other camera main control chip manufacturers will inevitably integrate the exclusive AI module suitable for security scenarios into the main chip in the future; the same, the security manufacturer headed by Haikang, Optimizing its own solutions, it is also expected to develop its own dedicated artificial intelligence chips in cooperation with related chip manufacturers. With the increasing demand for smart security capabilities in the society, the future ASIC master chip containing AI functions will become the mainstream chip of security.