news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Zhitong Finance APP learned that CITIC Construction Investment released a research report saying that computing power is still the fundamental driving force for the continuous iteration of large models. There are three main lines of investment in computing power. The first is around incremental changes, such as copper connections, liquid cooling, etc.; the second is around share changes, such as storage, PCB, power supply, etc.; the third is around whether Nvidia's performance growth rate exceeds expectations, which to some extent determines the valuation range of the entire computing power industry chain.

CICC pointed out that the trend of AI moving towards the end is inevitable. Apple's native AI operating system and applications created by Apple Intelligence have a certain leading position, and Microsoft's PC-side AI assistant is also accelerating. It focuses on the end-side replacement trend and investment opportunities such as upgraded DRAM, privacy computing, acoustics, batteries, heat dissipation, and Arm PC.

With the improvement of domestic large-scale model capabilities, the decline in call prices and policy support, CITIC Securities believes that more AI applications will be gradually implemented. Chatbots, image processing and video processing applications on the C-side are gradually being accepted. B-side AI is also beginning to be implemented in the fields of finance, industry, military, medical care, education, etc.

The main views of CITIC Construction Investment are as follows:

Looking ahead to the second half of the year for AI investment, we believe there are two major investment directions:The first is the global industrial trend, with computing power at its core; the second is to stimulate domestic demand based on domestic policies, especially the combination of AI with various industries on the G and B sides.

There are three aspects of investment logic in the computing power field: First, investment around incremental changes. The most important change in the second half of the year is the development and transition of AI server form from the past 8 cards to NVIDIA NVL36 and 72 cabinets. The cabinets are more integrated and are the main choice of many large manufacturers. Among them, copper connections and liquid cooling are new increments. The second half of this year will enter a period of intensive orders and performance catalysis will begin in Q4. Second, investment around share changes. With the spillover of orders from leading companies, the entire industry chain is highly prosperous, and the share of some companies has increased. Focus on storage, PCB, power supply and other sectors. Third, valuation fluctuations. The global valuation system of AI computing power refers to NVIDIA. Whether NVIDIA's performance growth rate exceeds expectations or not determines the valuation range of the entire computing power industry chain to some extent. Referring to TSMC Cowos expansion rhythm, it has accelerated in the second half of the year. We are optimistic about NVIDIA's performance in the second half of the year, so we are also optimistic about the entire sector's valuation switch to next year. At the same time, at the European Technology Seminar in April, TSMC announced that it would expand CoWoS production capacity at a compound annual growth rate (CAGR) of more than 60% until at least 2026. It can be seen that TSMC is optimistic about the overall demand for AI.

The value of copper connections has increased significantly and will increase in the second half of the year:1) The usage has increased significantly: Compared with the 8-card server that mainly used PCIE lines in the past, the NVL36/72 cabinet uses PCIE lines in addition to the computing trays. It also uses high-speed copper cables to connect between computing trays, between the switching chip and the backplane, and between the switching chip and the front panel I/O port. High-speed copper connections can also be used to interconnect different cabinets; 2) The industry chain began to increase in volume in the second half of the year: According to TrendForce, the new platform Blackwell will enter the market in Q3 and will begin to increase in volume in Q4. Looking ahead to next year, the Blackwell series will become NVIDIA's main sales product, and it will mainly be in the form of cabinets. For Chinese manufacturers, the main opportunity for copper connections comes from the spillover of Amphenol's orders. Considering the overall shipments of NVL36 and NVL72 cabinets next year (the equivalent NVL72 is expected to be 40,000 to 50,000 units), the market for high-speed copper cables in the cabinet alone will reach 4.8-6 billion. If the lines outside the cabinet are considered, the market size will be even larger.

The demand for cloud memory is booming:1) High demand in computing centers: The value of video memory accounts for a high proportion. The total demand for HBM3e and LPDDR5x in an NVL72 cabinet is about 207,400 US dollars, accounting for about 7% of the value of the NVL72 cabinet; 2) Demand for end-side equipment: In the long run, the parameters of end-side equipment will continue to increase, and the memory will continue to increase. The memory of the next generation of AI mobile phones is expected to grow to 12-16GB. The core investment opportunities in this industry chain are two points: 1) HBM3e share and changes in Apple's mobile phone 8GB DRAM and next-generation DRAM suppliers, with a focus on Micron; 2) At present, for large storage manufacturers, the main focus is still on production conversion. In the future, the balance of storage supply and demand may be broken, driving storage prices to continue to rise.

New upgrade of power supply solution:The server adopts a rack integrated design and uses an external unified power supply to further optimize power management and improve the overall performance and reliability of the system. NVL72 requires 6 power shelves, and a single cabinet requires a power supply of about 200kw. The DrMOS module is a specific power supply unit for the GPU, and the demand for DrMOS modules for the B series chips continues to increase. The main suppliers are Delta and Lite-On, and we will pay attention to the progress of mainland suppliers.

Comprehensive application of liquid cooling:At present, liquid cooling is mainly based on cooling plates and immersion. Compared with traditional air cooling, liquid cooling has three advantages in the context of increasing computing power density: 1) Liquid cooling cabinets have a higher power consumption limit and can carry AI servers above 20 KW; 2) The PUE value of liquid cooling rooms is closer to 1, meeting the latest policy requirements; 3) Under the background of low PUE, the same computing power demand consumes less electricity, and the long-term operating cost advantage is significant. Overall, the increase in the popularity of liquid cooling is an inevitable trend in the era of high computing power density.

Key investment trends on the AI ​​side: First, the increase in DRAM: For a model with 3 billion parameters, the memory usage is about 1.4GB calculated with Int 4 precision, and 6GB of DRAM is difficult to support large AI models. Currently, the only models that can support this large end-side model are those equipped with 8GB DRAM in Apple phones. We believe that if Apple upgrades its end-side model to 7 billion parameters in the future, the memory usage will reach 3GB, and the existing 8GB models will also be very difficult to support. In the future, each generation of mobile phone upgrades will be very important for Apple's DRAM upgrade. On the Android side, among Google's latest Pixel 8 series, only the Pixel 8 Pro supports running the AI ​​large model Gemini because it is equipped with 12GB DRAM. Second, security issues and privacy computing: In the future, AIPCs or AI phones will form local knowledge bases, and it is crucial to ensure the security of personal information. In the future, end-side security chips and algorithms will also be upgraded. In addition, Apple Intelligence will analyze whether the request sent by the user can be run on the device side. If stronger computing power is required, private cloud computing can be used to send only task-related data to servers using Apple chips. Some major mobile phone manufacturers will build their own cloud reasoning centers for mobile phone services in the future. The third is the upgrade of acoustics: Voice interaction will be an important entrance to the AI ​​era, and a very important point on the end side is the upgrade of acoustic devices. Fourth, the battery and heat dissipation have changed significantly: with the increase in the computing power of the end-side chip, power consumption has increased, the battery has become larger, and the heat dissipation material has also changed. Fifth, pay attention to the changes in demand for machine vision inspection equipment brought about by new changes in iPhone17 hardware. Sixth, Arm PC: With stronger AI capabilities, stronger battery life, Microsoft's key support, Qualcomm's efforts to build PC chips and other factors, major manufacturers have begun to focus on launching Arm PCs.

Tesla leads the end-to-end transformation of autonomous driving: Tesla FSD Beta V12.3 is the first FSD version to use an end-to-end neural network. According to Tesla's latest public safety data, after turning on the FSD function, Tesla vehicles may have an accident every 5.39 million miles, which is far lower than the average level of one accident every 670,000 miles in the United States. Compared with previous algorithms, large models have four important characteristics and advantages: 1) Data-driven: In the past, autonomous driving was defined by writing rules, but now it is data-driven. Every 1-1.5 million video clips are watched, the effect will be significantly improved; 2) High upper limit: The emergence ability of large models is currently also reflected in autonomous driving, which means that increasing the number of parameters can solve some difficult driving behaviors in the past autonomous driving process; 3) Faster iteration speed: In the past, Tesla iterated the FSD algorithm nearly every two weeks, and in the era of large models, the version was updated every 2-3 days; 4) Driving experience is close to humans: In the past, autonomous driving defined by rules was very rigid, and the current experience is more similar to the feeling of human driving, thereby reducing the number of takeovers. Tesla also released a series of data for FSD12.3, with the average takeover mileage increasing from 116 miles to 286 miles, and the proportion of trips without user takeover increasing from 47% in FSD V11.4 to 72%. The effect has been significantly improved, and the assisted driving capabilities of passenger cars have been further improved. Desay SV is a key focus. However, for L4 autonomous driving, Tesla's average takeover mileage is still far behind that of humans. At present, autonomous driving cannot be fully realized by relying solely on single-vehicle intelligence, and attention needs to be paid to the construction of domestic vehicle-road-cloud.

AI applications empower thousands of industries: Another main battlefield for the application of large models is industry applications. At the symposium of enterprises and experts in Shandong Province on May 26, the focus was on the use of technology to transform traditional industries. Secondly, ultra-long-term government bonds will also be invested heavily in areas such as "scientific and technological self-reliance". At the same time, the Outline of the Strategic Plan for Expanding Domestic Demand (2022-2035): Firmly implement the strategy of expanding domestic demand and cultivate a complete domestic demand system. It also focuses on promoting the deep integration of technologies such as 5G, artificial intelligence, and big data with transportation and logistics, energy, ecological and environmental protection, water conservancy, emergency response, and public services to help improve the governance capabilities of related industries. We believe that AI has begun to land in the fields of finance, industry, education, transportation, military, and medical care. On the financial side, large models have gradually become better investment research assistants, wealth management virtual people, financial knowledge bases, etc. On the industrial side, large models have begun to provide human-computer interaction, AIGC sample generation, etc. in software such as CAD, focusing on central control technology. In the field of robotics, the intelligence of robots connected to large models has increased rapidly, and they have begun to replace people to complete simple tasks in factories and other scenes. In the military field, the overseas company Palantir has successfully used big models as battlefield assistants. In the field of education, AI has gradually become a virtual teacher in more subjects. In the field of transportation, vehicle-road-cloud collaboration puts higher demands on infrastructure, enabling intelligent traffic management while effectively reducing the cost of intelligent driving vehicles. In the medical field, AI itself has been deeply applied in medical imaging, new drug research and development and other fields in the past (traditional models). The emergence of generative models has further deepened the development of AI in the above fields, but overall, overseas research and development directions are more inclined to pharmaceuticals, and domestic research and development directions are more inclined to health management. There are differences in the application directions of the two based on the effectiveness of big models.

risk warning

Expectations for economic recession in North America are gradually increasing, there is great uncertainty in the macro-environment, and changes in the international environment affect the supply chain and overseas expansion; the chip shortage may affect the normal production and delivery of related companies, and the company's shipments are lower than expected; demand and capital expenditures in informatization and digitalization are lower than expected; intensified market competition has led to a rapid decline in gross profit margin; rising prices of major raw materials have led to lower-than-expected gross profit margin; exchange rate fluctuations have affected the exchange gains and gross profit margins of export-oriented companies; the update and iteration effects of large model algorithms are lower than expected, which may affect the evolution and expansion of large models, and in turn affect their commercialization; progress in automotive and industrial intelligence has been slower than expected, etc.