XXX, 您好 登录 注册 登出
首页 / 新闻中心 / 新闻详情

中国国际电视台CGTN连线驭数高级副总裁张宇总

发布日期:2022-04-08

张宇总就未来算力应用、DPU需求等方面内容进行专业讲解。

采访视频链接:https://yusur.tech/zkls/zkys/video_detail.html?id=71&v=0.4202167493469866

以下为采访实录。

1.Help us understand the critical role that computing power plays in China's future economic growth... How important is it to the digital economy era?

Thanks, and my pleasure to be on the show. Computing Power is the foundation and the most important cornerstone of the digital economy. Actually, there won't be a "digital economy" without Computing Power. As stated in the report, for US and China or any countries whose computing power index score is over 60, if the Computing Power Index grows by 1 point, we would expect 3 times relative growth in GDP. According to another Computing Power report from CAICT (the China Academy of Information and Communication Technology institute), 1 RMB invested in computing power building, will result in 3 to 4 RMB in total GDP. So it is indeed very important to both digital economy as well as the whole national economic growth. Computing Power is basically influencing almost all the areas of our life and economy, from online to offline, from industry to financial, from game to education or research and etc. So it is very important.

 

2.Quantum computing is believed to be the next highly disruptive technology in computing power. How are China and the US positioned in this race towards this exponentially higher speeds than today’s most powerful supercomputers?

Well, Quantum Computing is to harness the unique properties of quantum mechanics to solve certain types of problems far faster than traditional computers or supercomputers. Given time, it will be the major source of future computing power. However, Quantum Computing is still in its infancy, and it will still take quite some time for the Quantum Computing era to come. Quantum technologies are developing very fast, though. Globally we could see all the major countries, including China, are increasing their investments in Quantum Computing, and the total amount adds up to more than 13 Billion US dollars.

As to position, we do see the US is taking the lead in terms of papers published or cited and patents filed in Quantum Computing, but China is ranked second and catching up very fast. At the same time, if you look at the Quantum Communication area, China is actually leading in terms of paper publishes.

So I would be very optimistic that China will become a major contributor or even leader in the areas of Quantum Information Technology. Still a lot of hard work to do, though.

 

3.Does your company do anything in particular to facilitate the utilization of the (computing) power? 

Yes, Yusur is focusing on DPU (data processing unit) design and development with our own Domain-Specific Architecture. Despite being relatively new, DPU is becoming one of the key processors to make data centers more power-efficient and help meet the Computing Power need of processing even larger amounts of data. Actually, DPU is a typical milestone of Domain-Specific Architecture development. As we all know, processors like CPU have been the main sources of computing power. And much of the computing power improvement in processors comes from decades of miniaturization of semiconductors, a trend that is called Moore's Law. Basically, in 1975, Intel founder Gordon Moore predicted the regularity of the semiconductor miniaturization trend, which doubled the number of transistors on computer chips or processors every two years. That means exponential growth in computing power lasted for the past few decades. However, unfortunately, Moore's Law has been running out of steam as a viable way to grow computer power recently. By 2018 it showed a roughly 15-fold gap between Moore's Law prediction and current capability in performance. So comes the post Moore's Law era, and in this era, processors created with the Domain-Specific Architecture method are more efficient than traditional processors in different domains and together with CPU, could continuously provide more computing power with much less power consumption. Actually, we have also seen this trend in the popularity of GPU, AI Chips, and Bitcoin Miner processors. Looking forward, DPU will be a key driver to revolutionizing the current computational infrastructure of data centers. It is positioned as the third computational processor together with CPU and GPU in data centers by the industry.

 

4.The global data processing unit market is experiencing a significant growth. What trends will drive the future demand of DPUs?

Well, DPU could process IO-centric and Data-centric tasks very efficiently compared with CPUs, such as virtualized network, virtualized storage, securities computation and etc. Looking forward, it is a must for large amounts of data processing in data centers. There are two major trends behind DPU's popularity. First, for the last few years, we've witnessed exponential growth in the amount of data to be processed in the digital economy globally. According to the IDC data, we are expecting 163ZB of data in 2025 and the data growth rate is actually catching up each year, especially in the new areas like autonomous driving, 5G and Edge Computing, the Industrial Internet, VR/AR or even Metaverse (the data are increasing in different orders of magnitude). Combined with the post-Moore era situation, we need DSA processors like DPU or AI chips to largely increase computing power in different domains to process these data efficiently. At the same time, with so much data produced every day, the network bandwidth to move these data around also need to be increased exponentially. These are the IO-intense as well as data-intense tasks mainly, and DPU is order-of-magnitude more efficient than CPU or GPU to process these kinds of tasks. Moving forward, we believe, together with CPU, GPU and DPU, we will have a much more balanced yet efficient computing infrastructure.

 

5.What does it mean for the grid network, telecommunications network and also resource consumption? 

With Industrial internet and edge computing gradually rolled out, they are posing tough challenges to the network as well as computing power. It's not just about having enough computing power. It also requires the network to be extremely fast with low latency and high throughput. For example, the Industrial Robot Control requires less than three milliseconds of end-to-end latency. This couldn't be done with the previous telecommunication network, which couldn't handle it fast and stable enough. But with 5G, it is achievable. Actually, these kinds of requirements are the major user scenarios targeted by 5G. China has the leading position worldwide in terms of 5G construction and has the most 5G base stations installed. At the end of 2020, there have been around 770,000 5G base stations deployed in China. Meanwhile, in order to meet all these new strict requirements, the convergence of computing and networking is also taking place, where the computing power will be allocated and transmitted with a holistic view of the overall tasks. This way, computing power and related resources will be consumed more efficiently than before. This is called "Computing Power Network", and all the major telecommunication providers in China are pushing hard on building this.

 

6.Your company is one of many which are actually involved in building such infrastructure. Could you share with us what kind of infrastructure are you building and why is it important?

Yusur is focusing on DPU development. As I explained previously, DPU will be the critical processor to handle the IO-Intense and Data-Intense tasks efficiently. Currently, for large data centers, 25G is the standard bandwidth of the network interface for servers in those data centers. Without DPU, processing IO and virtualization at a speed of 25G has already consumed up to 30% to 40% of the CPU's processing power. And with an exponential data growth rate, we are expecting the standard bandwidth in large data centers to be increased from 25G to 100G within the next 4 or 5 years. When it comes, CPU only won't be able to handle so much data at such a high speed. So basically, without DPU, all the computing power of CPU will be consumed just by moving data around, packaging and unpacking data, stuff like that. There won't be extra CPU power to actually process the data or complete algorithms like AI for meaning. Thus, it is only with DPU in the near future that our large data centers could continuously serve the digital economy well with exponentially-growing data volume. Worldwide, all the major cloud service providers, such as Amazon, Microsoft, Alibaba, Tencent and etc., have deployed or been deploying DPU to their Data Centers.

 

7.From the perspective of private sector, what potential and opportunity will the new project unleash?

The project will unleash huge opportunities for computing power equipment providers or processor designers like YUSUR. For example, currently, there are around 5 million Server Racks deployed totally in data centers in China, and by the end of 2025, this number will become 18 million. It's more than three times growth within the next five years. So it is a significant business opportunity for all related companies or start-ups.

At the same time, since many new data centers will be built from scratch,, in order to achieve Carbon Neutrality, these new data centers will most likely be equipped with the latest technologies and types of equipment, like DPU. This will greatly boost the new technology adaptation rate in data centers and help those companies or start-ups who have been focusing on tech innovations related to computer power infrastructure. I do see great opportunities and potential.

 

8.By better utilizing renewable energy in the west, the project plays key role in lowering the carbon footprints of computing. What more can be done to cut CO2 emissions? 

From a Computing efficiency point of view, I do think continuously increasing the tech innovation speed, and adaptation rate would help, especially for the legacy data centers with legacy equipment and network, where the PUE (Power Usage Effectiveness) is not good. How to revamp them step-by-step with new architecture and equipment would also be important to cut overall carbon CO2 emissions.