Audio and video architecture construction in the super video era ｜ science and Intel jointly launched the second season of "architect growth plan"

By the end of 2022 year , The number of Internet video users in China has reached 9 Hundreds of millions of people ;IDC（ China Internet data center ） expect ,2025 In the global data space 80.3% Will be based on video 、 picture 、 Unstructured data represented by audio , The efficient processing of audio and video is becoming more and more important .

The rise of short video and live broadcast has expanded new business forms , Live video 、 Intelligent content generation 、 More and more applications such as video recommendation , It has brought huge development opportunities to the video cloud business , At the same time, it also poses new challenges to the cloud service architecture , High concurrency is involved 、 Distributed storage 、 Audio and video coding and decoding, edge computing and other technologies .

Young architects are interested in architectural design 、 Technology selection is often full of doubts , from Science Jointly launched with Intel “ Architect growth plan ” The second quarter series of courses is tailored to such needs , Carefully selected audio and video 、 Cloud games 、5G Core network 、 Computational networks 、 Federal learning 、 Ten hot topics, such as big data of Shengxin , Provide high-quality learning resources and practical experience sharing for architects .

To help architecture technicians in the audio and video industry answer questions ,“ Architect growth plan ” Tan Dai, general manager of volcanic engine, was specially invited in the first phase 、 Cheng congchao, senior chief engineer of Intel big data 、 Liu Jiang, vice president of Zhiyuan Artificial Intelligence Research Institute, jointly discussed 《 Construction and evolution of audio and video architecture in the super video era 》.

Volcano engine ： Audio and video architecture in the super video era

The first lecturer Tan will combine the specific practice of volcanic engine , Focus on the technology direction of video cloud backup ： Edge of computing 、 Audio and video codec 、 The application of intelligent midrange technology , Introduce the evolution of audio and video architecture in the super video era . The course is divided into three parts ：

The first part ： What is the super video era

2020 year , The outbreak , All walks of life or actively or passively start to contact with video , The video cloud penetrates more “ Tradition ” industry , In the industrial 、 education 、 More and more landing scenes are emerging in medical treatment , These scenarios bring functions to the audio and video architecture 、 Different challenges at the performance and security levels , The video cloud has come to the super video era .

The second part ： Design of ByteDance audio and video service architecture

The edge is closest to the business scenario , First, select the rich edge resources and operator networks in all provinces and cities in China , And deploy high-quality single line according to the geographical level 、 Multiline sum BGP The node of , Hardware equipment combining multiple architectures , Such as ：X86、ARM The server 、 Smart network card 、GPU etc. , Realize the edge foundation base for heterogeneous computing force , Build from 1ms To 40ms Wide area network access and edge data processing capability . The cloud native edge platform is built based on the edge infrastructure base , Flexible management of heterogeneous computing and network resources , Realize one network scheduling at the edge .

In order to reduce the super large scale 、 Real time processing challenges the central architecture , Based on ROI Video coding concept , To structure the corresponding computing and storage architecture , This not only benefits from bandwidth costs , In terms of user metrics , Including the average duration, there was also a significant increase . Through self-developed multi scene data sets 、 Eye tracker acquisition ROI, And then in CPU Upper use mobilenet Accelerate time domain modeling ,GPU Parallel processing on , Realized >90% The accuracy of , Video compression performance is also significantly improved .

Optimize the experience through data driving , Build a complete QoS、QoE Data systems , And continue to optimize , Re evolve with data-driven technology . From the collection of data 、 mining 、 Model training to strategy distribution , To tandem AB The feedback of the experimental platform is all open , In different users 、 The optimization can be personalized in different scenarios , Refine , Maximize cost performance .

The third part ： The future evolution trend of video cloud and what kind of video technology it needs

For the super video era , Video cloud technology should better meet the user's immersive 、 Interactivity 、 The ultimate HD video experience . The lecturer combined with the large-scale... In this year's Beijing Winter Olympic Games 8K UHD broadcast , It is always a great challenge to propose video coding and transmission , Need top-notch video compression capability , The end cloud integration of volcanic engine H.266 Video coding scheme , Saved 30%-50% Bit rate of , Pave the way for Ultra HD video .

Intel shares ： End to end video optimization integrating hardware and software

Cheng congchao, senior chief engineer of Intel big data , In the first session, I brought you “ An end-to-end video optimization scheme integrating hardware and software ”.

Focus from input to output , From software to hardware , Production from content 、 Storage 、 Calculation and distribution , Interpreted the full stack optimization scheme of Intel video cloud .

In video coding and decoding , Intel has developed its own software based on CPU The codec component of the processor Scalable Video Technology（SVT）,SVT Technology to realize the block parallel of a graph 、 Parallelism between graphs 、 Parallelism between multiple frames of video , Through several layers of parallelism, the whole CPU Make full use of all the cores of , In each core, it passes through AVX-512 and SIMD Optimization of instruction set , In a CPU Do more processing in the instruction cycle of , Final realization 2-20 Times the video codec performance .

SVT With a highly scalable Core Architecture , Fully realize SIMD/AVX-512 Optimization of instruction set , be based on IntelXeon CPU The platform achieves better thread and process concurrency , Take full advantage of multi-core functionality , Video transcoding speed reached 、 Video quality 、 The best tradeoff between transmission speed .

similar SVT Such software layer optimization and underlying infrastructure XPU After the interface of is encapsulated , Depending on the load , Can pass OneAPI To call different processing units at the bottom , Where does the data execute , Can automatically sense and schedule , Make full use of the cloud 、 Border 、 Terminal processing capacity , Realize encoding and decoding 、 Reasoning 、 Maximize the efficiency of rendering .

Dialogue between big names ： How does the video Cloud Architecture balance the software and hardware investment

AC collision spark , Communication inspires . In this round table dialogue , Liu Jiang, vice president of Zhiyuan Artificial Intelligence Research Institute 、 Lecturer Tan Dai 、 Lecturer chengcongchao and three famous professors surrounded “ Construction and evolution of audio and video architecture in the super video era ”, Conducted in-depth exchange and discussion .

Liu Jiang ： In the era of live and short video , How the volcanic engine gets through AI、 Cloud technology to improve the application experience ？

Tan Dai ： Integrate rich edge nodes and network resources in the world , Whether it is traditional audio and video applications , Or a new edge computing scenario , Can get a quick response , Specific to audio and video , Achieve a balance between performance and experience through efficient encoding and decoding technology ; And set up a system based on QoS and QoE The index system of , Continuously improve the user experience with data drive .

Liu Jiang ： When the processor reaches the bottleneck , How to improve the processing efficiency of audio and video ？ What solutions does Intel have for audio and video ？

Cheng congchao ： Intel has done a lot ” Soft and hard integration ” Industry solutions . In short, it can be summarized into three big pieces —— Hardware aspect , Increase speed and reduce cost , The cloud computing power in the future must be XPU（ Multiprocessing unit ） Solutions for ,CPU、GPU、IPU A variety of combinations , More flexible distributed computing power ;DPDK、SPDK The network transmission is greatly optimized ; Software aspect , Intel has been committed to contributing to the open source community , We do upstream/downstream, Try to enable the open source community . At the industry level , Make industry solutions with partners .

XPU As an important cloud computing power , What are the new breakthroughs in deep learning reasoning ？ How to build the architecture of video recommendation technology , How to support 100 million level video content understanding and distribution ？ The iteration cycle of hardware is longer than that of software , As an architect, how to balance the investment of software and hardware resources , Achieve the highest cost performance ？ Welcome to watch more wonderful content “ Architect growth plan ” Complete course .

Course link ：

• https://bizwebcast.intel.cn/eventstart.aspx?eid=298&tc=9m8w007tzd&frm=wechatkol