With the increase of the number of high-definition surveillance cameras and the increase of AI popularity, the images and video stream data collected by cameras need more powerful analysis engines to analyze, process and train.
Take Beijing subway station as an example. There are more than 65,438+10,000 subway stations in Beijing, and each station has hundreds of cameras. It is common for each subway station to circulate 80 to 65,438+10,000 people every day. It is conservatively estimated that each camera watches 654.38+00000 people every day, and then assuming that there are 654.38+00000 targets in the comparison database (not big for the public security database), the question that this camera has to answer every day is 1.00 1 billion!
Obviously, in all kinds of current security projects, relying on simple embedded intelligent DVR and NVR can no longer meet the strict calculation requirements.
Facing the trillion-dollar AI security market, at the crossroads of technology, all security enterprise executives will face an ultimate proposition: What kind of powerful server does AI security need?
When it comes to secure servers, X86 is ubiquitous and has always been the darling of many industries, including security.
"At present, many management platforms in the security market, such as streaming media servers, forwarding servers and master servers, are basically based on X86 architecture. Its biggest advantage is that it is relatively easy to develop and quick to get started. Most engineers are better at research and development on X86 architecture. "
Fu Jianhui, CEO of Huatai Jack, told Lei Feng. Considering its good cost performance, X86 servers have always been our first choice for purchasing.
Therefore, X86 servers have been favored by the world's top server suppliers for many years.
Unfortunately, X86 servers with "up to standard" scores in all subjects have encountered some "local problems" in the face of today's hot AI wave.
Fu Jianhui revealed that from the actual use of security users, there are three main problems in the application of X86 servers in the security industry at present:
First, CPU is not responsible for many logical operations, so it is inefficient in multitasking. In the face of massive video information, the traditional X86 server deployed in data center with CPU as the core can no longer meet the requirements of parallel flexible computing and changeable environment, and it is difficult to make an amazing performance in the server market of security enterprises.
"The previous video data only needs to exist in the background, and doing a small amount of analysis means that the storage capacity is large enough; Today, many customers hope that we can process these massive video information in real time and feed back the results, which means that the system needs to do decoding, video structuring, recognition, search and so on. At the same time, X86 is obviously not enough. "
In other words, X86 can be compared to a function machine in a mobile phone, which can meet a single communication processing requirement. In the security market where AI is integrated, a more powerful smart phone is needed to meet the personalized needs such as games and image processing.
Secondly, under the background of insufficient computing power in the industry, many manufacturers have made a combination of X86 servers and several GPU cards. This "patchwork" scheme that simply solves the computing power greatly increases the power consumption of servers and the cost of users.
From the perspective of industry adoption, when it comes to AI projects such as face recognition, most manufacturers will use GPU as the structured processing unit of portrait data, especially in X86 server clusters, GPU is the only choice.
To some extent, GPU does solve some problems of insufficient computing power, but it also has two fatal injuries.
First, it consumes a lot of power and needs to rely on X86 architecture servers to run, which is not suitable for a wider range of AI solution development; Second, the cost is high. For example, using GPU scheme, the cost of converting one-way face recognition is more than 10 thousand yuan. Compared with other schemes of 1,000 yuan or even 100 yuan, there is no cost advantage. These two fatal shortcomings also make many enterprises have to seek new solutions.
Thirdly, because X86 uses more open LinuX systems than closed AIX systems, it is slightly insufficient in stability and maintainability.
"The future market must be the competition of data scale and computing power."
Zhang Qi, the product department of Inspur Commercial Machinery Co., Ltd., told Lei Feng. Com believes that with the emergence of more and more new applications, the traditional X86 computing architecture will encounter many bottlenecks, including data bottleneck (the speed at which the computing unit of the processor can acquire and exchange data), computing bottleneck (how much computing power can be integrated per unit space), delay bottleneck and communication bottleneck.
Just as a road with a design speed of 30 yards per hour is difficult to carry vehicles with an average speed of 100 yards, it can cause road congestion or even paralysis in a short time.
Today, in the face of big computing and intelligent scenes, whoever can solve the problem of computing power first and better reduce power consumption and cost will be the leader in the AI wave.
In Zhang Qi's view, a high-performance server based on POWER9 can meet the high-intelligence requirements of the AI ? ? security era.
According to the actual situation of AI security, Inspur Commercial Machine Co., Ltd. recently launched an AI vision analysis intelligent analysis solution (UltraVision on Power) based on POWER9 server and equipped with UltraVision video intelligent analysis system.
AI vision analysis solution can be regarded as an ultra-efficient AI brain, which combines software and hardware, and can complete complex data processing required by various industries including security in real time, accurately, intelligently and energy-saving.
"Hard" is embodied in the architecture of POWER9, which can provide powerful computing and processing capabilities for images and videos. Compared with other processors, POWER9 supports the next generation I/O protocols, such as PCIe4.0 and NVlink2.0, which can show better application performance in AI and other applications.
Specifically, compared with X86, the number of single-node video processing paths increased by nearly 3 times, reaching 3.8 times, which improved the training efficiency of AI model of deep learning framework, accelerated the database performance by 1.8 times, and improved the IO capability by nearly 5 times.
In addition, when performing video and image coding and decoding, query and search tasks, the whole machine can provide super-computing power with single precision of 56TFlops and double precision of 28TFlops. Compared with X86 server, a single GPU can provide 30 times higher reasoning ability than a pure CPU server.
It is worth mentioning that the unique CAPI technology of this scheme can reduce the delay to 1/36, accelerate the image processing in an all-round way, and reduce the power consumption by up to 30%.
1.8 times, 3.8 times, 3 times, 5 times, 30 times, seemingly small numbers, are huge orders of magnitude for the security industry.
The change of these figures can advance the intervention of prevention and control means of various crimes and serious violent incidents before or during the event, greatly reduce the occurrence of public security incidents, and realize the business transformation of public security from passive defense to active defense.
In addition to the super computing power provided by POWER9, on the software level, the scheme also includes high-weight UltraVision video intelligent analysis technology, such as target detection (PD), pedestrian re-identification (RE-ID) and many other computer vision technologies, which improves the accuracy of target recognition as high as 94%.
Undoubtedly, the AI vision analysis solution combining software and hardware can solve the problems of high computing power and low power consumption in the AI era for users in the actual landing process.
In addition, compared with other popular schemes, this scheme has two major advantages.
First of all, using the unique ability of old resources can reduce customer costs.
Generally speaking, if a general AI video system wants to realize some functions, it must be connected with a perceptual camera with AI technology. In the deployment process, the solution does not need to replace the original camera, and only needs to bypass the video acquisition terminal to realize the AI system.
In addition, the scheme can also be compatible with cameras of any different brands and standards; It can be directly deployed without changing the hardware architecture of the customer's original server, which effectively reduces the deployment cost of the customer.
Second, as far as Inspur Commercial Machine Company itself is concerned, relying on its leading position in the server field, it has a strong customization landing capability and shortens the delivery cycle from months to days.
Whether facing AI applications such as big data processing and machine learning, or open source applications such as software-defined storage and in-memory database, this scheme will have a good performance.
Undoubtedly, the artificial intelligence visual analysis solution specially designed for emerging applications such as artificial intelligence, cloud computing and big data provides customers with more diversified choices when facing severe business challenges.
Relying on this high-performance product, users can deploy all kinds of intelligent applications faster and shorten the technical iteration cycle of secure AI applications.
At the same time, the application of Inspur commercial server with excellent performance is not limited to the security industry, but also can be used in areas with high security requirements such as the Internet and finance.
In addition to security, the whole society is transforming and upgrading to scale, automation and intelligence. Among them, the intelligent application direction covers four major directions: front-end, cloud, platform and industry.
In this upgrading process, the new platform needs to have new capabilities to do new cognition, and new cognition gives birth to new demands and applications.
For technology companies including Inspur, this is a great opportunity and a great challenge. It's a long way to go, Xiu Yuan, and you must go up and down. Leifeng net Leifeng net Leifeng net