CCTV Beijing, July 4 (Reporter Qi Zhiying, Intern Xie Rongrong, Li Qingyi) What is the current situation and prospects of the construction of computing infrastructure in the mainland? What are the opportunities and challenges for future development? How to coordinate the development of the upstream and downstream of the industrial chain? How to reduce power consumption in the face of high power costs? ……
The eighth issue of "Shanghai Market Hui Hard Science and Technology Hard Guest" focuses on "building a computing power base", which was recently recorded and launched, and the guests of 4 leading enterprises on the computing power supply side of the Science and Technology Innovation Board face the hot topics in the industry with professional thinking and enterprise development experience, and jointly explore the high-quality development path of computing infrastructure construction in mainland China.
The guests who participated in the recording included Sha Chaoqun, General Manager of Haiguang Information, Wang Yang, Executive Vice President and General Manager of Business Operations Department of VeriSilicon, He Han, CEO of BIWIN Storage, and Yang Lianfeng, Director and President of Primarius Electronics. Kuai Jian, deputy director of the Oriental Securities Research Institute and chief analyst of the electronics industry, served as the leading guest of this program.
"Cutting-edge science and technology, hard customer forward-looking"! The column aims to build a space for in-depth exchanges between the iconic souls of the development of "hard technology", so that entrepreneurial scientists and entrepreneurs can summarize their experiences, review their paths, look forward to the industry and make suggestions, so as to further lead and help the continuous improvement of the ecology of the industrial chain of the Science and Technology Innovation Board, and then achieve high-quality development.
The recording site of "Shanghai Market Hui Hard Science and Technology".
What factors accelerate the increase in the scale of computing power in mainland China?
According to data from the Ministry of Industry and Information Technology, by the end of 2023, the total scale of data center racks in use in China exceeded 8.1 million standard racks, and the total scale of computing power reached 230 EFLOPS, ranking second in the world.
In October 2023, the Ministry of Industry and Information Technology and other six departments issued the "Action Plan for the High-quality Development of Computing Infrastructure", which provides target guidance for the construction of computing power, carrying capacity, storage capacity and application empowerment in the mainland. In terms of computing power, by 2025, the scale of computing power will exceed 300 EFLOPS, and the proportion of intelligent computing power will reach 35%.
Sha Chaoqun, general manager of Haiguang Information
"2023 is an important turning point in the development of artificial intelligence, and it is also the year of the explosion of global computing power scale and services. From the perspective of computing power types, general computing power, intelligent computing power and supercomputing computing power are increasingly showing a trend of integration, and in the future, general computing power with super-intelligent integration will become the main demand of the market. Sha Chaoqun said.
Wang Yang, Executive Vice President and General Manager of Business Operations Department of VeriSilicon
When it comes to the pattern of the global computing power market, Wang Yang believes that 2024 will show significant regional differences and technology-driven diversification trends.
He Han, CEO of BIWIN Storage
He Han is full of confidence in the development of the mainland's computing industry, and believes that the mainland's computing power construction is in a stage of gaining momentum and is about to usher in a new era of vigorous development.
Yang Lianfeng, Director and President of Primarius Electronics
"Enterprises in the computing power industry chain should have both confidence and determination to promote the development of the country's computing power strategy together." Yang Lianfeng emphasized that computing power is not only a market development demand, but also a national strategy.
Based on the views of the five guests, the scale of computing power in mainland China can accelerate its growth, thanks to the joint promotion of policy guidance, technological innovation, industrial application and infrastructure construction.
At the policy level, a series of favorable policies, including Eastern Data and Western Computing, and the Action Plan for the High-quality Development of Computing Infrastructure, have been promulgated, clarifying development goals and key tasks, and promoting the high-quality development of computing infrastructure.
At the technical level, with the development of artificial intelligence, big data and other technologies, the demand for computing power continues to increase, which promotes the innovation and improvement of computing technology.
At the industry level, the digital transformation of all walks of life, especially in the fields of government affairs, industry, transportation, and healthcare, has a growing demand for computing power.
At the infrastructure level, the construction and upgrading of computing infrastructure such as national eastern and western computing nodes, data centers, intelligent computing centers, and advanced computing centers in various regions have provided a physical foundation for the growth of computing power.
At the market demand level, mainland artificial intelligence technology is accelerating into the era of comprehensive application, and large models and generative artificial intelligence are deeply empowering all walks of life, which has given rise to a huge demand for computing power and promoted the rapid growth of computing power scale.
What is the importance of the coordinated development of the upstream and downstream of the industrial chain?
Kuai Jian, deputy director of the Oriental Securities Research Institute and chief analyst of the electronics industry
Kuaijian believes that in the computing power industry chain, enterprises in different links need to cooperate together to make the industry bigger and stronger.
The four companies attach great importance to ecological construction, Haiguang Information has established the Haiguang Industrial Ecological Cooperation Organization (hereinafter referred to as "Photosynthetic Organization"), VeriSilicon has organized the Qingcheng Mountain Forum and Dishui Lake Forum, etc., and Primarius Electronics, MPI, Rohde and Schwarz have jointly hosted the "Building a Semiconductor Test Ecosystem" Semiconductor Electrical Testing User Conference.
"'Go fast alone, go far together', the development of domestic chips and computing power is inseparable from an open and collaborative industrial chain." Sha Chaoqun said that Haiguang Information has invested great energy and manpower and material resources in ecological construction, with the aim of uniting ecological partners to jointly build an open and win-win whole industry chain from domestic high-end chip design to application services.
According to Sha Chaoqun, there are currently more than 3,000 partners in the photosynthetic organization, and Haiguang Information and ecological partners are in a relationship of mutual trust, equality, mutual benefit and win-win, and uphold an open attitude towards partners.
Wang Yang also affirmed the importance of the coordinated development of the upstream and downstream of the industrial chain for the entire semiconductor industry. He believes that the coordinated development of the upstream and downstream of the industrial chain will help optimize the allocation of resources, reduce production costs, and improve the efficiency and competitiveness of the entire industrial chain.
"Through collaboration, on the one hand, upstream and downstream enterprises can share resources, including technology, information, market channels, etc., which will help reduce R&D and production costs and accelerate product innovation; On the other hand, it can also promote the exchange of knowledge and technology, and promote the innovation of the entire industry. Wang Yang said.
VeriSilicon attaches more importance to the technical capabilities of its partners, the spirit of continuous innovation, market reputation, strategic consistency, consistency of development strategies of both parties, stability and reliability, and responsiveness.
He Han introduced that BIWIN Storage not only actively participates in the construction of a domestic ecosystem in the storage industry, but also initiates the establishment of the Shenzhen Memory Industry Association, which serves as a platform to promote exchanges and cooperation in related fields such as industrial control, vehicle-grade storage, server storage, etc., and gathers industry forces to promote the integration of multiple resources through organizing forums and matchmaking meetings.
At the same time, He Han also said that the inherent volatility of the storage industry has led to a high degree of concentration in all links of the industrial chain, and mutual trust and cooperation need to go through a long-term and multi-cycle test to jointly cope with the ups and downs of the industry. BIWIN Storage attaches more importance to the capabilities of partners in terms of technology complementarity and innovation, supply chain stability and quality control, market collaboration and brand influence.
"The integrated circuit industry started from the business model of the foundry in the 90s, and the upstream and downstream industrial chain has formed a highly coordinated global ecology, and EDA is one of the main supporting links." Yang Lianfeng believes that from the perspective of the industrial chain behind the computing power, the improvement of the computing power of the chip alone is not enough to meet the processing needs of AI.
Nvidia CEO Jensen Huang has said that in the next 10 years, Nvidia will increase the computing power of deep learning by another 1 million times. Where does this 1 million times come from? It must require innovation in systems, interconnection, architecture, and algorithms, combined with the continuous evolution of models, which can only be achieved through industrial chain collaboration. Yang Lianfeng said.
How to promote the process of products from the laboratory to the market?
As a technology-intensive and talent-intensive industry, the chip semiconductor field on the supply side of computing power is mostly characterized by large investment, long cycle and slow return. How do companies move from the lab to the market?
In 2023, Haiguang Information will invest 2.81 billion yuan in R&D, accounting for more than 40% of operating income, R&D technicians will account for 92% of the total number of employees, and nearly eighty percent will have a master's degree or above. Driven by R&D, Haiguang Shuangxin has been widely used in data centers in telecommunications, finance, Internet and other industries, as well as in big data processing, artificial intelligence, business computing and other fields.
Sha Chaoqun said that Haiguang Information's advantages in two aspects have promoted its process from the laboratory to the market. First, Haiguang Information adheres to the product research and development strategy of "sales generation, verification generation, and R&D generation", and the launch of products has a reasonable planning and market rhythm.
Second, Haiguang Information has a strong advantage in ecological partners, and there are more than 3,000 ecological partners in the "Photosynthetic Organization" to empower each other, which also helps Haiguang Information to bring products to the market more smoothly and efficiently.
VeriSilicon's technology investment has also achieved good results. According to IPnest data, in 2022, VeriSilicon's IP licensing business market share ranked first in China and seventh in the world. In 2022, VeriSilicon's IP licensing revenue ranked fifth in the world. According to IPnest's IP classification and public information of each company, VeriSilicon's IP types rank among the top 20 IP companies in the world.
It is worth noting that according to the revenue calculation in 2023, 67% of VeriSilicon's revenue comes from the one-stop chip customization service business, and 33% comes from the semiconductor IP licensing service business.
"Through the synergy of the two businesses, the two-legged walk is more stable and more competitive than simply taking IP licensing or design services." Wang Yang shared VeriSilicon's strategy of building high competition barriers and forming a technological moat.
In 2023, BIWIN Storage will make breakthroughs in the research and development of core technologies such as IC design, advanced packaging and testing, and chip testing equipment research and development. With the support of R&D and technical strength, the company achieved revenue of 1.727 billion yuan in the first quarter of 2024, a year-on-year increase of 305.8%, and a net profit attributable to the parent company of 168 million yuan, a year-on-year increase of 232.97%, a record high since listing.
Regarding BIWIN's advantages in promoting from the laboratory to the market, He Han believes that it lies in the practice of two core strategies. First, BIWIN Storage has its own market-oriented genes, and will do what it can to ensure the stability of profitability and cash flow at each stage of the entire industrial chain, and the accumulated resources are continuously invested in basic capacity building, so as to realize the upgrading and upgrading of the industry step by step, forming a virtuous circle of self-reinforcement; Second, BIWIN Storage is committed to building differentiated competitiveness, taking the lead in building a business model of integrated R&D, packaging and testing, and built its own packaging and testing plant in 2010.
Primarius has also continued to invest heavily in R&D in recent years, with R&D investment accounting for more than 72% of operating revenue in 2023, and a high proportion of R&D investment has also injected growth momentum into the company's rapid development, with a compound growth rate of more than 40% in the past three years.
"Focusing on the strategic layout of technology and product co-optimization of process and design, the company continues to develop, evolve and expand core technologies, and has taken the lead in making breakthroughs in the technology in the two fields of device modeling and circuit simulation. According to Yang Lianfeng, the innovative research and development of core technologies and the transformation of achievements are the key foundation and driving force for the company to continue to maintain international competitiveness.
At the same time, Yang Lianfeng also affirmed the importance of the help of the Science and Technology Innovation Board and the cultivation of talents.
What is the accumulation of technology in the direction of reducing power consumption?
According to the China Academy of Information and Communications Technology, the cost of electricity in data centers accounts for 60%-70% of their total operating costs. With the continuous growth of computing infrastructure and the popularization of technologies requiring higher computing power density such as artificial intelligence, the energy consumption of computing infrastructure itself will also bring considerable challenges.
Kuai Jian said that the entire industrial chain is contributing to energy conservation and emission reduction, and companies are improving computing power on the one hand, and fully considering power consumption on the other hand.
"Energy savings in data centers are becoming more and more important now." Sha Chaoqun pointed out that Haiguang Information reduces power consumption from three levels.
One is the CPU and DCU chip itself, Haiguang Information has a series of low-power design technologies, and at the same time has complex power consumption management for different application scenarios; The second is to improve the efficiency of the server, Haiguang Information works with system manufacturers to optimize the software and improve the load efficiency to reduce power consumption; The third is to reduce the PUE (the ratio of all energy consumed by the data center to the energy consumed by the IT load), and Haiguang Information's state-of-the-art immersion cooling can reduce the PUE to 1.04, thereby reducing the power consumption of the entire data center by about 27%.
"For data centers and computing infrastructure, the cost of electricity accounts for a high proportion, and we must not talk about computing power regardless of energy consumption." Wang Yang said that VeriSilicon's high-performance and low-power technology has been widely used in data centers and servers.
According to Wang Yang, VeriSilicon's first-generation video transcoding acceleration solution can achieve 6 times the transcoding capacity of traditional high-end CPUs, consume only 1/13 of its power consumption, and have a very small chip area, which can help cloud service providers greatly reduce the power consumption and floor space of servers. The second-generation product has greater optimization in terms of power consumption and performance; On this basis, VeriSilicon has also customized 5nm media accelerator card chips for data centers for leading international chip companies, which have now entered the mass production stage.
"First at the main control level, the first is the architecture, we have multi-level sleep and multi-level clocks to do low-power design; Second, in terms of process, we try to choose some low-power libraries; Thirdly, in terms of algorithms, we can reduce the number of migrations and energy consumption by improving the error correction capability of LDPC. He Han shared three ways to reduce power consumption by BIWIN Storage.
"In the field of manufacturing, in addition to the manufacturing process of the fab, we can also reduce power consumption through packaging and testing technology." He Han said.
Opportunities and challenges for the expansion of AI computing clusters
With the growth of large model parameters, the scale of AI computing clusters is getting larger and larger, and NVIDIA recently announced that it will launch Blackwell Ultra and Rubin GPUs next year, respectively, and the corresponding Ethernet network platforms are Spectrum-X800 and X1600, which can form clusters with a scale of hundreds of thousands and millions of cards.
What opportunities and challenges may the increasing number of clusters bring to enterprises on the supply side of computing power?
Sha Chaoqun believes that the construction of ultra-large-scale clusters is accelerating, which puts forward higher requirements for processors, storage and network technologies, and at the same time brings opportunities for technological innovation and product upgrades.
"Chip design manufacturers, machine manufacturers, software teams, and application parties...... We must optimize and improve together to create an easy-to-use, easy-to-use, and reliable system to be able to cope with large models with more than one trillion parameters in the future. Sha Chaoqun emphasized.
From the perspective of the industry trend of cluster development, in the face of future challenges, Wang Yang said that it is not easy to do a good job in every link, but there is an opportunity to do a good job, and we need to be cautious and optimistic, step by step.
"When we are catching up, we also have to find some special advantages of our own to pull up, so that we can speed up the pace of catching up. In this process, we need to consider not only the IP, but also many details in the chip design. For example, the size of the chip area, computing power, memory mode, packaging form, as well as the cascade mode, the drive mode, the choice of interconnection after multiple cards, and the energy consumption, system-level design, etc. Wang Yang pointed out.
In the field of storage and packaging, He Han believes that after the construction of the Vanka cluster or even a larger computing cluster, first of all, the scheduling of memory will face a big bottleneck, and the industry's idea to solve this problem is to do memory pooling.
"Intel's launch of the CXL protocol is also to prepare for the pooling of memory, CXL is currently in the 2.0 industry promotion stage, may be 3.0 is an explosive period, will quickly adapt to the general trend of Vanka cluster construction. Therefore, we are also actively investing in the CXL field, in the 2.0 generation, we are now doing solutions and modules, and in 3.0 we will discuss whether to intervene in the design of the main control and the entire software and hardware ecosystem. He Han said.
In addition, He Han believes that the second challenge of interconnection will be interconnected after the increase of network cards and computing cards, and there are different levels of interconnection to solve and improve efficiency.
"Artificial intelligence and large models may bring some opportunities to the chip industry in the future, and may make some new possibilities for chip design and manufacturing." Yang Lianfeng believes that EDA may be a carrier of these possibilities.
"The research and development of EDA tools itself is a process of using a large number of algorithms to optimize and generate relevant content through the learning and recognition of existing data, and it should be said that EDA is a relatively early field using AI technology." Yang Lianfeng said that Primarius is also doing some exploration in the field of artificial intelligence. (CCTV Capital Eye)
For more exciting information, please download the "Central Radio Network" client in the application market. Welcome to provide news clues, 24-hour hotline 400-800-0088; Consumers can also complain online through the "Woodpecker Consumer Complaint Platform" of CCTV. Copyright Notice: The copyright of this article belongs to CCTV, and it may not be reproduced without authorization. Please contact [email protected] for reprinting, we will be held accountable for not respecting the original.