In the era of super video, what kind of technology will become the base?

“2 For an hour and a half , I was so nervous that I was sweating .” Lihuiqin talked about the first live broadcast of goods , I couldn't help laughing . As the first batch of professional technicians to minning town , During the epidemic, Tiktok, which had not been played for several years, resumed , Opening “ Mayor Li praised minning ”, Live delivery , Introduce minning town 、 Promote Ningxia specialty . Plus TV series 《 Mountains and seas 》 Hot broadcast , Since then, the people of the whole country have known that Fujian （ Fujian ） And Ningxia （ ning ） The immigrant town established for poverty alleviation , Mayor Li's fans have increased rapidly , Lycium barbarum in Ningxia 、 The mushroom 、 Beach mutton , It was also sold from mayor Li's live studio to the whole country .

The popularity of the Internet , The establishment of information superhighway , Let small mobile phones become the property of more and more farmers “ New farm tools ”, Visible specialty goods , Under the traction of invisible data torrent , Heading for the whole country 、 To the world ,“ Cloud economy ” Become a key contributor to poverty eradication 、 A powerful starting point for Rural Revitalization and common prosperity .

As a new type of infrastructure to promote China's high-quality development , Upgrading of computing power and bandwidth , For more and more huge audio 、 Video streaming provides an unobstructed highway , The Internet under the development of digital technology such as streaming media + Cloud technology , To industry 、 Social development has released great potential and vitality , Become the way to win under the new infrastructure of enterprise digital transformation .

1. Decoding new infrastructure

30 Years of Internet development have brought earth shaking changes to China , It is becoming one of the important driving forces to promote China's economic development . The digital economy continues to grow rapidly , Accelerate the integration of information technology and the real economy , The new momentum of development brought about by it gradually grows , It has become a new engine of China's economic growth . In the process , from 2013 to 2020 year , China's mobile Internet access traffic is from 13 Billion GB Growth to 1656 Billion GB, The compound annual growth rate is as high as 99.86%. Due to the 2021 year 6 month , China Mobile Internet access traffic is 1033 Billion GB, Year-on-year growth 38.6%.

Google predicts , The Internet of the future 80% The content presentation form of will be video . Visual information replaces tedious 、 Lengthy 、 Obscure words 、 Data or symbols , It can more intuitively reflect the essence of information , More directly promote the exchange of information between the two sides .

Issued by China Internet Information Center The first 48 Time 《 Statistical report on Internet development in China 》 Show , By 2021 year 6 month , The number of short video users in China is 8.88 Billion , The average usage time of short video applications per person per day has exceeded 120 minute .

QuestMobile data display ,2020 In, the scale of Internet advertising in China reached 5439.3 Billion , On the ability of media form to seize advertising , Video information flow grows faster , Has become a mainstream form of advertising , Gradually replace the share of graphic information flow , Typical advertisers have significantly increased their budget for short video media . According to the prediction of Zhongyan Puhua , Short video platform TikTok Advertising revenue may be in 2022 Tripled in , achieve 110 Over US $100 million , Surpass its competitors Twitter And Snapchat The sum of the two .

From live delivery to short video publicity , Then to the increasingly popular distance education and mobile office , Video or wealth password of value mining , Or the highway that delivers value , High quality and low-cost video delivery capability , Become one of the core competitiveness of enterprises , It also shortens the distance between people , Human society has thus entered a new stage of development .

Behind this , It is the full support of the new network infrastructure .2018 year , The central economic work conference defined for the first time “ New infrastructure ”, At the top of the list is information infrastructure , Include 5G、 The Internet of things 、 Industrial Internet 、 Communication network infrastructure represented by Satellite Internet , With artificial intelligence 、 Cloud computing 、 New technology infrastructure represented by blockchain , Data center 、 Intelligent Computing Center is the representative of computing infrastructure .

2. Build audio and video capabilities

In conjunction with volcanic engines IDC released 《 Evolution trend of video cloud in Super Video Era 》 The white paper points out that , Today's era has experienced long video 、 Short video and live broadcast application stage , Enter the super video era . In this era , Internet users have a higher demand for video , High definition 、 Interactivity 、 Immersive experience has more demands . Video cloud is also accelerating its penetration into traditional industries , In the industrial 、 education 、 More landing scenes are emerging in traditional fields such as medical treatment , This brings functionality to the audio and video architecture 、 Different challenges in performance and security . among , According to Intel data , Now big data 90% The above comes from unstructured data , Mainly videos and pictures , The rapid growth of the number of videos and users has brought great pressure on bandwidth and performance , How to optimize is a big topic .

2022 year 2 month , Country “ Count from east to west ” The project is fully started , Start the construction of a national computing hub node , And planned 10 A cluster of national data centers . In the eastern data calculated by applying the Western computing power resources , The proportion of video data cannot be underestimated . These data come from many heterogeneous monitoring platforms , The number is huge and there are different standards . This requires computing infrastructure , It must be able to support the storage of massive video data 、 analysis 、 Indexes 、 retrieval 、 transcoding 、 Application etc. , Avoid video storage and application in different areas and links , And can integrate different devices 、 Different platforms 、 Gather video resources of different standards , Realize effective sharing and integrated utilization .

meanwhile , machine learning 、 Computer vision 、 Natural language processing, etc AI Applications are also emerging , These all mean frequent access to memory and data transmission , This breaks through the traditional von Neumann computer architecture , As a computing infrastructure, data center and Intelligent Computing Center , Targeted solutions are needed .

The evolving data center needs a new intelligent architecture . In Intel China Cloud & According to Cheng congchao, senior chief engineer of big data , This intelligent data center architecture will have three types of computing units —— For general purpose computing CPU、 Acceleration for a specific application or workload XPU And for infrastructure acceleration IPU—— They will be connected to each other through a programmable network , So as to make effective use of data center resources . In especial , Facing the performance intensive load in diversified media applications , Super computing power and the ability to accelerate the encryption, decryption and transcoding process , It can greatly improve image processing 、 The efficiency of AI applications such as image recognition and video analysis .

if X86 CPU Built Intel's glorious past ,IPU Will become the key for Intel to open the future world .IPU It is a programmable network device , It aims to reduce the number of cloud and communication service providers in the central processing unit （CPU） The cost of this , Release valuable CPU Kernel resources . utilize IPU, Customers can deploy secure, stable and programmable solutions , So as to make better use of resources , Balance the workload of data processing and storage , So as to greatly improve the revenue of the data center .GPU It greatly improves the processing ability of the application program for two-dimensional and three-dimensional graphics and images . Besides , In processing intensive data and parallel data computing ,GPU It can provide dozens or even hundreds of times more than CPU Performance of , Therefore, in the field of large-scale parallel computing applications , Such as light and shadow processing 、 Financial analysis 、 Medical imaging 、 Oil and gas exploration 、 Geographic Information 、 Genetic analysis and scientific calculation, etc , Begin to play an increasingly important role . More importantly, with AI Appearance ,GPU Suitable for large-scale parallel floating-point operations 、 The characteristics of matrix calculation have been brought into full play .

But whether it's CPU still GPU, It is obviously unrealistic to achieve higher computing efficiency with a single architecture . use CPU+GPU Heterogeneous computing mode （ Include CPU+ Coprocessor 、 Speed up the processor, etc ）, It's going to be a trend .

3. Password of traffic

stay 2021 Year of CVPR On , from AI Benchmark Host 「 Move AI challenge round 」 Draw attention to , And in it 「 Real time mobile terminal detection scene competition 」 The team that won the championship , The result is more than eight times higher than the second place .

This competition requires the use of mobile terminal deep learning framework TensorFlow Lite CoreML Run real-time mobile scene detection , requirement AI The algorithm predicts the scene category of the image input by the camera in real time , share 30 Scene categories , The training set is only 9900 A picture , For AI , There is a great risk of over fitting .

From byte skipping ByteScene The team used the transfer learning method to come up with a capacity of only 8.2Mb The little model of the , Realized 95% Of Top 1 Recognition accuracy , The average delay is only 4.4 millisecond .

Now? , This real-time mobile scene detection technology has been applied in products such as Tiktok and clipping . Because the operation is simple , Intelligent use experience , This technology improves App User experience on Internet . This is also the technical capability that volcanic engines are exporting .

Volcano engine is the cloud service platform of byte beat , Relying on Cloud Computing 、 big data 、 Artificial intelligence and other technical capabilities , Provide technical products and solutions for customers , Including infrastructure 、 Cloud native 、 Personalized recommendation 、 Audio and video processing capabilities, etc. in Tiktok 、 Today's headline 、 Technology implemented in watermelon video and other applications .

As video cloud services enter the super video era , Internet users have a more extreme demand for video , No longer satisfied with a clear and smooth viewing experience , But for high definition 、 Interactivity 、 Immersive experience has more demands . meanwhile , The penetration of video cloud into traditional industries is accelerating , In the industrial 、 education 、 Medical and other fields , More diversified landing scenes have emerged , These scenes also bring new functions to the audio and video architecture 、 Different challenges at the performance and security levels .

As the volume of data to be processed continues to rise , Audio and video applications have increasingly stringent requirements for the performance of computing and storage architecture . Especially in high throughput 、 In the scenario of large data requests , We need to find a solution that can expand the capacity economically and efficiently without affecting the system performance .

Volcanic engine is in the process of helping enterprises apply , An innovative path has been explored in the framework —— Through the integrated architecture of end cloud , From the support of edge infrastructure 、 Calculation + The application solutions from network digital basic services to video cloud provide production 、 Storage 、 Full stack capability from processing to distribution .

As Tan Dai, general manager of volcanic engine, said ：“ If the platform architecture is not easy to use , There is no way to realize video cloud services for user experience . On top of edge infrastructure , Video cloud also needs a lot of encoding and decoding 、AI Analysis and other algorithm processing , How to carry the demand of super large-scale real-time processing , This brings many challenges to the central architecture .”

Although by expanding memory , It can realize the performance advantage of enhancing delay and throughput , But the capacity that this solution can expand is always limited , and DRAM The high price of memory and the cost of more rack space , It will also greatly push up TCO. So , Volcano engine uses Intel aoteng persistent memory and Intel Xeon scalable processor to deeply optimize the storage architecture of content information business , It can break through the limitations of storage media and architecture , It can also achieve the gain of input and output on the premise of ensuring performance .

Intel aoteng persistent memory is not only close to DRAM Memory performance , There is also SSD Data persistence for , It can make up for DRAM Memory and SSD Performance gap between ; It's relative to DRAM Large capacity of memory 、 Low cost features , It can also help enterprises achieve a better balance between capacity and cost .

Volcano engine is using Intel aoten persistent memory App Direct In mode , Persistent memory is directly connected to the processor through the memory bus , Applications can bypass the operating system , No driver required 、 system call 、 Interrupt and context switch can directly access the data on persistent memory from user space , therefore , The system has little software overhead , The delay is about 100ns-300ns, Not only better than SSD, It is also close DRAM Memory performance , It can well meet the demand for data access speed during peak periods .

This model also has data persistence . Whether encountering unexpected power failure or planned shutdown , The data in the persistent memory can be preserved for a long time , Thus, storage reliability can be improved 、 Increase the business flexibility of the system 、 Shorten the system restart time and improve the fault recovery speed of the recommended system .

meanwhile , In order to meet the demand of Ultra HD , Transcoding performance is the key . As a new generation of video codec standard ,H.265/HEVC（HighEfficiencyVideoCoding, Efficient video coding ） It can still ensure high-definition video quality under high compression ratio , Favored by major live video platforms , But because of the complexity of its algorithm and data structure , The complexity of encoding and decoding 4 Twice as many as the previous generation H.264 codecs , Therefore, the performance requirements of the processor performing transcoding work , It's very prominent .

Intel Xeon scalable processor , Provide performance support for the live video broadcast process including transcoding . meanwhile , Intel Advanced vector extensions built into Intel Xeon scalable processors 512（ Intel AVX-512） It further strengthens its performance . The powerful transcoding performance ensures that the video is H.264、wma、rmvb、avi Fast conversion between other formats , And through flexibility 、 Compact player , Help live broadcast operators through mobile apps 、 Web pages and other display methods quickly and comprehensively cover the audience .

4. Towards a new era

For applications , The key is to experience . In the age of Super Video , Higher definition 、 More interactive 、 The demand for more immersive video technology is changing with each passing day , The optimization of technology can make more interactive gameplay and business scenarios possible .2021 year 10 month , Mark, CEO of Facebook, an American social platform · Zuckerberg announced , Facebook changed its name to “ element ”, In the future, the company will give priority to the related businesses of yuancosmos . Immersive experiences such as augmented reality and virtual reality , Bring the meta universe into life , Help people connect 、 Develop business , Build the next development of social technology .

The basic features of the metauniverse include ： Immersive experience 、 Virtualization separation 、 Open creation . To realize the meta universe , Need a series of IT Comprehensive application of new technology , Including network and Computing Technology （5G、6G、 The Internet of things 、 Cloud computing 、 Fog calculation and edge calculation ）、 Management Technology （ Energy management 、 Resource management 、 Session management and space-time consistency management ）、 Virtual and real object connection 、 Modeling and management technology （X Connected to the Internet 、 Identity modeling 、 Social Computing 、 Decentralized management technology ）、 Virtual and real space interaction and fusion technology （ Expand the reality 、 Video game technology 、 Brain computer interface technology ） etc. .

Meta universe 、 Immersive experience , And large-scale AI deployment , It also puts forward requirements for future graphics processing products , Need to achieve a high degree of visual fidelity 、 Optimized AI and continuous computing . This brings strong new demand for data centers in the cloud era , New architecture is needed , Realize intelligent computing 、 Massive data processing 、 Integration and acceleration of heterogeneous computing power 、 Multi load application support .

In Cheng congchao's view , For virtual reality / Augmented reality , The underlying infrastructure is needed to support , The front-end device can play , The network can transmit , Backstage data center , Can be right VR Enhanced rendering batch 、 Real time and efficient processing . So , Intel has built a robust accelerated computing system and graphics card roadmap , Across all products from the client to the data center , Help customers cope with different needs in the digital age .

For the future data center , Intel believes that there will be four features ： Realize the physical separation and logical integration of computing and storage , Realize the integration of security features in the internal architecture of the chip , Achieve more flexible hardware 、 Software 、 Applications and services , The software will be presented in the form of microservices , Realize ubiquitous heterogeneous computing .

“ Computing everywhere ”, It's Intel CEO Pat · Kissinger put forward “ The human digital society has four super technological forces ” One of , And ubiquitous connection 、 From cloud to edge infrastructure and artificial intelligence , Help mankind continue to grow , Continuous innovation , Full empowerment “ Everything is intelligent ” The transformation of .