At present, both users, traffic and video resolution are constantly increasing. Compared with voD services that require less data processing and sufficient processing time, real-time video services that require higher data processing volume and more stringent delay control will bring great challenges to basic codec servers. For enterprises, it also means that cost control and profit generation will be more and more pressure. This post was posted on LiveVideoStackCon2019 Beijing by Sean Gardner of Celings DCG Data Center.



The text/Sean Gardner

Organizing/LiveVideoStack

Hello, my name is Sean Gardner, who is in charge of the Video business of Celings DCG Data Center. The main content of this speech is that Celings seeks new business exploration and technical innovation in the development process of China.


1. Xilinx application fields and challenges


The spirit has been widely used in every field, there is no lack of among them a lot of innovation, it is worth mentioning the spirit thought has great exploration and practice in 5 g, especially in the cloud based applications and is sensitive to real-time application of two fields, one of the most representative are cloud games and is sensitive to network latency of video conference system (VSaaS).



The chart above highlights the importance of real-time video and communications in all markets, and specifically addresses the importance of customer original video and content, which creates infrastructure challenges and opportunities.


Regardless of the segment, the industry is currently facing two major challenges: the increasing cost of hardware and operating costs, including operating data center hardware and server prices are becoming more expensive.


2. Current situation and challenges of live broadcasting and voD industry


It can be found from the market research results that the two applications of live broadcast and on-demand will face a great range of growth in the future. The market predicts that the demand for CDN of live broadcast business will reach $70 billion in 2021, and may exceed that of on-demand business in 2023.



As can be seen from the figure below, mobile video live broadcasting will occupy an increasingly large market, and China will become the largest market for mobile video.



In addition to the booming Chinese market, the US has two major streaming and on-demand apps: NETFLIX and Twitch. NETFLIX is mainly a file-on-demand service. The file traffic processed in actual application scenarios is far less than the content in live broadcasts, and the on-demand service can have enough time to process video files. Twitch mainly provides game live streaming services. In terms of servers, the infrastructure has to process a larger amount of streaming videos to achieve real-time or even millisecond processing, which poses greater challenges to servers and requires higher costs in all aspects.



The curves above show the large differences in peak bandwidth, peak traffic, and average traffic, which Silas will provide more efficient management and solutions for servers.



In the above table, except the Bandwidth % of Revenue is calculated according to certain data, all other data are public and objective. As mentioned before, the cost of server and Bandwidth becomes very expensive, which can be seen from this table, especially why the cost of real-time video traffic is rising sharply.



Bandwidth usage has two major challenges: OPEX and CAPEX. At both extremes, Ninja had 600, 000 simultaneous viewers, compared to Four for Alfred. The 80/20 rule mentioned here is that 20% of video streams account for 80% of network traffic, and the other 80% account for 20% of network traffic.


Use Twitch data to illustrate the 80/20 rule: 13% of video streams account for 74% of web traffic


In addition to the increase in bandwidth costs, the cost of all infrastructure is getting higher and higher. The red curve in the figure below represents the increase in cost, and the yellow and green curve represents the ARPU. You can see the flattening of profitability, which brings another challenge — how to make money.




In addition, as mentioned before, the number of users and traffic is increasing, and the figure below shows that the resolution of video is also increasing to 8K or even higher during this period. At the same time, video encoders are becoming more and more complex, with coding standards ranging from H.264 to H.265 to HEVC/VP9 and AV1, and coding tools are becoming more and more complex, which poses a huge challenge for infrastructure servers.



3. DCG video soft IP policy High VQ/low bit rate policy



According to the previous summary, the left side is the development of software business based on very traditional FPGA, and the right side is the development of new hardware business. As you can see from the figure above, the left side is a huge source of network traffic, with a particularly large number of viewers. While the right side occupies a lot of video streams, for example, security video has a lot of traffic, but not many people watch it. Celings offers different solutions for these two scenarios: eSports and SafeCity are two completely different video scenarios, so we offer two completely different solutions based on their differences — high quality on one side and high density on the other.


OPEX requires developing software solutions, such as Intel’s, based on traditional hardware. In a software solution based on traditional hardware, you have to choose between quality and speed. The yellow line on the right must sacrifice video quality for real-time video speed.



In contrast, Silas added Soft IP on the basis of hardware, the quality can be compared with better video encoders. At the same time, due to the adoption of the underlying hardware scheme, the speed is far beyond the software scheme.


Xilinx and X265 bit rates are compared under the same PSNR


Comparison of video coding rates with the same quality


Take Tigress Live as an example, the bandwidth cost of tigress live increased by 66.8 percent per quarter, and the bandwidth cost of tigress live increased by 25 million DOLLARS per quarter. Syrynx can save 40 percent of the bit rate, which is 40 percent of the bandwidth, on the same quality, and 25 million dollars a quarter in bandwidth costs if you can save 30 or 40 percent that means you can save over 21 million dollars a year in bandwidth costs.



In order to achieve bandwidth savings for users, Selings itself and the entire ecosystem of partners provide various video and image transcoding tools, including different coding standards, encoding formats, and different container formats.


Xilinx Ecology provides a variety of tools for bandwidth saving services


It is worth mentioning that Selings also creatively integrated FFmpeg on Selings FPGA, as long as you are familiar with FFmpeg, you do not need to have a high level of understanding of the underlying FPGA. All command lines are the same. The only difference is that there is an option in them.


4. Innovative video solutions


4.1 Socionext uses Xilinx FPGA in the next generation solution


Socionext has developed a very high quality H.264 encoder on a XILINX ALVEO card and performed both subjective and objective tests on the encoder. The following figure shows the test results. Objective quality assessment includes VMAF and PSNR. All the red lines in the figure show the performance of SocionextH.264 Encoder.


Socionext H.264 Enc. Vs. X264 Enc. Very Slow Setting


In addition to objective evaluation such as PSNR and VMAF,FPGA H.264 also has obvious advantages in subjective evaluation. As can be seen from the comparison below, THE H.264 encoder based on FPGA of Selinth has more details and has obvious subjective advantages in subjective quality.


Comparison between FPGA H.264 Enc and Nvidia T4 H.264 Enc
4.2 V-NOVA encoder encoding efficiency



Another IP on The Syrnes FPGA is the V-Nova encoder. Currently, the V-Nova encoder can achieve resolution of 1080P, and in terms of encoding efficiency, it can compress the bit rate of 1080P video to 950K, which is exciting in many countries or regions where the bandwidth environment is not ideal.
4.3 Xilinx has great room for improvement in the future



Because fpGas are programmable, there is a lot of room for improvement in both density and video quality. It is important to note that both OPEX and CAPEX, both high density and high quality, are based on the same FPGA acceleration framework and have exactly the same user interface.



Finally, Celings and partners across the ecosystem work together to create two different solutions for different scenarios, as well as hardware solutions and software systems on top of hardware.