2024-08-12
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
Original first release | Jinjiao Finance (ID: F-Jinjiao)
Author | Jiaoye
The storm is coming.
Just when the price war among car companies was in full swing, a bucket of cold water was poured on them. Car companies have received information about the self-sufficiency rate of chips in order to prepare for the extreme pressure that may be put on them one year later. At present, the situation is extremely pessimistic. The self-sufficiency rate of automotive chips is less than 10%, and domestically produced chips are far from being able to lead the market.
In contrast, Nvidia's "special edition" chips continue to sell well in China. Although the computing power performance has been castrated by 85%, Nvidia predicts that it will export 1 million AI chips to China this year, with a total value of about 90 billion yuan.
The only hope is still Huawei, but the reality is far more complicated than imagined, and Huawei is not a panacea.
A self-inspection storm is sweeping the automotive industry.
A number of automakers have received notifications requiring them to complete a survey of their chip self-sufficiency rates before the end of September.Foreign media reported that the Ministry of Industry and Information Technology required BYD, Geely and other electric vehicle companies to expand the purchase of domestic electronic components and accelerate the use of domestic semiconductor chips. It is reported that there was an informal goal that required automakers to increase the purchase of domestic chips to one-fifth by 2025, but now the pace of the promotion of the proportion of domestic chips is very unsatisfactory.
Subsequently, a report by experts on domestic substitution of automotive chips circulated in the market, confirming"Reach the 25% localization rate red line in 2024." For enterprises that fail to reach the red line, preferential policies will be gradually reduced, and there will be mandatory requirements for state-owned enterprises.
What is more noteworthy is that in this minutes, the data on "automotive chip localization rate" is sorted out in detail, which is shocking:
The localization rate of SoC chips, the "brain of the car", is currently less than 10%.The main suppliers are still Nvidia, Qualcomm, Intel, etc. Domestic suppliers such as Horizon Robotics and Black Sesame have a very low market share.
The localization rate of MCU chips, the "automotive nervous system", is less than 10%.There is almost no localization in the high-end MCU field, and the main suppliers are NXP, Infineon, Renesas, etc. Domestic suppliers such as BYD and Xintangwei have a certain share in the mid- and low-end markets.
In the field of "automotive memory" storage, the localization rate is also less than 10%.The main suppliers are Micron, Samsung, Hynix, etc. Domestic suppliers such as Yangtze Memory Technologies Co., Ltd. and Changxin Memory Technologies Co., Ltd. have performed poorly in the market.
"Automobile muscle" power semiconductors have a relatively good localization rate, especially IGBT and silicon carbide, the localization rate reached 30% and 35% respectively. However, the localization rate in the main drive field is relatively low, about 20%. Major domestic suppliers include CRRC Times Electric, BYD, etc.
In the field of sensors for the "five senses of automobiles", the localization rate variesThe localization rate of some types of sensors, such as temperature sensors, is relatively high, reaching 60%-70%. However, the localization rate of pressure sensors and acceleration sensors in the mid-to-high-end market is relatively low. Major international suppliers include Honeywell, Infineon, etc.
In simple terms,In the field of mid-to-high-end automotive chips, the localization rate is very pessimistic.Luo Daojun, senior vice president of the Components and Materials Research Institute of the Fifth Institute of Electronics of the Ministry of Industry and Information Technology, publicly stated:
"China has the largest production capacity for new energy vehicles, and the use of (chips) is also increasing.The self-sufficiency rate of chips is currently less than 10%, it is a structural shortage.”
It can be seen that the price war among automakers is in full swing, but the fate of the core is still in the hands of outsiders. Some analysts also pointed out that even in the automotive chip market with low technical barriers and high localization rate,Relying on price wars, considering the overall cost of use, local companies may not be able to defeat international manufacturers.
For example, chip manufacturers must do sufficient quality control and function verification to reduce the concerns of small and medium-sized companies about replacing chips. Only with excellent quality can truly achieve benign domestic substitution. Otherwise, it can only become a negative example for international manufacturers to prove the excellence of their products.
As the US election situation changes, more severe extreme pressure may soon come. Some industry insiders call for"Independent brands give up internal war and quickly solve the problem of being stuck in chips"BYD, Great Wall, Geely, and SAIC should stop looking down on each other. When being strangled by others, if anyone has key components, can they share them with other domestic brother companies?
The question is who has such hard power? Some enthusiastic netizens pointed out that Huawei can basically make the above chips.It is also suggested that Huawei sell 5 million fewer low-end nova phones and transfer chip production capacity to domestic automakers., the most difficult problem of car-computer chips has been solved. Huawei has produced 2 million smart driving chips in full, which can meet the needs of high-end cars.
This is easier said than done.
The plot in which Huawei forced Nvidia to lower its prices has recently seen a reversal.
According to relevant media reports, Nvidia has lowered the price of its H20 artificial intelligence chip for the Chinese market. In some cases, the price of the H20 chip is more than 10% lower than that of Huawei's Ascend 910B. The reason is that the Ascend 910B chip performs better than the H20 in some key indicators, which has forced Nvidia to cut prices due to poor sales of the H20.
But now,It is difficult for Huawei to maintain its advantage.A report from Morgan Stanley pointed out that Nvidia's H20 series of artificial intelligence chips specially supplied to the Chinese market has begun to attract purchasing interest from Chinese technology giants including Baidu, Alibaba, Tencent and ByteDance.
The Financial Times of the United Kingdom even broke the news: Nvidia predicts that it will export 1 million AI chips to China this year, with a total estimated value of about 9 billion pounds, equivalent to about 90 billion yuan. At the same time, chip consulting agency SemiAnalysis estimates that Huawei's top AI chip Ascend 910B will sell about half of this number this year.
From the performance point of view, Nvidia's H20 chip, which is specially supplied to the Chinese market, is more accurately called a "castrated version of H100 computing power". Its AI computing power is only 15% of H100, and some of its performance is indeed inferior to Huawei's Ascend 910B. In this case, why do Chinese companies still spend 90 billion to buy it?
One theory is that although the H20's "paper" capabilities are lower than Huawei's Ascend 910B chip,But in actual use, Nvidia's H20 chip is "clearly ahead", thanks to the H20's superior memory performance.
Another reason is thatHuawei chips cannot compete with Nvidia in terms of software ecosystem.The switching cost is relatively high for purchasing companies, and in order to fully utilize Huawei's Ascend AI chip to meet the needs of large model training, they currently must rely on the assistance of the Huawei team. The migration work may take 9 to 12 months depending on the complexity of the large model, and engineer training will take at least 3 to 5 months.
A typical case is the "Spark All-in-One" device released by iFlytek. In order to successfully install the Huawei Ascend 910B chip, Huawei spared no manpower cost and deployed hundreds of engineers to help iFlytek adjust the parameters.
However, the above defects are not completely unacceptable.The more critical reason is that Huawei’s chip production capacity has been hindered.
Due to well-known reasons, TSMC is unable to manufacture Huawei chips, and Huawei is unable to purchase the equipment and spare parts needed for chip production. Huawei's existing 7nm chips and the 5nm chips that have reportedly been taped out are all produced using DUV lithography machines using multiple exposure technology. Although this solution is effective, it has a low yield and causes high losses to related equipment.
This has led to a foreseeable serious bottleneck in production capacity for Huawei. Due to the extended chip delivery time, market rumors have said that AI development at Alibaba, Tencent, Baidu and other companies has been impacted. They originally expected to use Huawei's chips to reduce their dependence on Nvidia, but now it seems that they can only turn around and buy Nvidia's H20.
The above prediction shows that Huawei can produce 500,000 AI chips this year, which can basically meet users' orders, which is quite optimistic.
In addition, since Huawei itself sells both water and goods, and is deeply involved in the fields of automobiles, mobile phones, and large models, it is inevitable that its competitors will worry about the issue of "selling their souls."
Therefore, although Huawei has the ability to design chips and partially manufacture chips, it is obviously unrealistic to let Huawei make all chips, which will only harm Huawei.
The automotive chip and AI chip markets are undoubtedly vast and boundless.
In terms of quantity, a single fuel vehicle only needs 300 to 400 chips, while new energy vehicles and vehicles with assisted driving functions will need more than 1,000 chips. Cars that achieve fully autonomous driving will use more than 3,000 chips.
According to IC Insight, by 2030, the global demand for automotive chips will exceed 100 billion, with the Chinese market alone requiring 46 billion. In terms of cost, the chip cost of a traditional fuel vehicle is about 2,270 yuan, while the chip cost of a new energy vehicle is about 4,540 yuan, twice that of a traditional fuel vehicle.
Industry insiders analyzed thatChina is rapidly advancing smart driving and driverless driving, essentially hoping to take the lead in power semiconductors and automotive smart chips.Because the Chinese market is so large, once China has a first-mover advantage, it will be irreversible. With the high-end development of automobiles, the chip war is inevitable.
In this regard, relevant national departments have also taken action. The Ministry of Industry and Information Technology recently issued the "Guidelines for the Construction of the National Automotive Chip Standard System (2024 Edition)" and proposed:
By 2025, more than 30 key automotive chip standards will be formulated, covering key product and application technology requirements such as control chips, computing chips, and storage chips, as well as vehicle and key system matching test methods;
By 2030, more than 70 automotive chip-related standards will be formulated, ensuring that key areas such as basic and general requirements are supported by standards, accelerating the healthy development of automotive chip technology and products.
In the view of Cui Dongshu, secretary-general of the National Passenger Car Market Information Joint Conference, this policy orientation will promote the development of domestic chips, accelerate the pace of domestic chips "entering vehicles", and is also one of the important measures to solve the "bottleneck" problem of chips.
On the other hand, AI chips are also experiencing new changes.
The general large models represented by ChatGPT have a bottomless demand for computing power and data.Most chip manufacturers have begun to look for more realistic business paths, which currently focus on large vertical application models in specific fields or industries.
For example, in the field of autonomous driving, we can focus on developing dedicated large models for this field, and only need to process nationwide traffic data to meet the needs; in the field of food delivery and express delivery, dedicated AI large model robots can be customized according to specific scenarios.
Compared with general large models such as ChatGPT that pursue large and comprehensive, vertical large models focus on small and precise, thus reducing the requirements for performance such as computing power. From this perspective, Chinese companies' large-scale purchases of castrated H20 chips may be based on the change in their judgment of AI trends.
Judging from the latest trends in the industry, this change is spreading rapidly. Except for a few giants such as Huawei that continue to compete with Nvidia, most chip companies are turning to the implementation of large models/small models in various industries, starting with inference scenarios that do not require high hardware and software, and starting to focus on niche markets. For example, they focus on low-power small chips for some power-sensitive scenarios; or they focus on niche scenarios such as video optimization to do small but beautiful business.
In addition, there are rumors that Huawei cannot dominate the domestic chip industry and requires a certain market share to be given to other chip companies such as Haiguang and Cambrian. According to the latest meeting spirit, capable private enterprises are supported to take the lead in undertaking major national technological research tasks.
It can only be said that relying solely on Huawei to solve the chip bottleneck is far from enough.
Collect and organize the latest news information and make it available for free for everyone to read
