After Musk's visit to China, Tesla's stock price soared, and Musk's value soared to $192.3 billion, returning to the world's richest man.
Nvidia's astonishing skyrocketing wealth has made it impossible to hold back, and according to insiders, Nvidia CEO Jensen Huang will visit China in early June.
Market sources said: Huang Jenxun will meet with executives of Tencent, ByteDance, BYD, Xiaomi, Ideal and other enterprises during his visit to China. The purpose is naturally to sell his AI chips.
With the explosion of ChatGPT, the AI industry has developed rapidly, and the demand for computing power as the foundation and foundation of AI has skyrocketed, reaching 300,000 times that of 10 years ago. This makes AI chip leader NVIDIA make a lot of money.
Affected by this, on May 30, NVIDIA's stock price soared to $419, with a market value of more than $1 trillion, becoming the world's "most valuable" chip company.
In order to ensure its own advantages and further expand the global market, NVIDIA continues to show goodwill to China, and CEO Huang Jenxun bluntly said "don't underestimate the strength of Chinese chips."
So the question is, will NVIDIA continue to provide AI chips for Chinese companies? Is the domestic AI chip bought or independently developed?
Where is the bottom strength of the NVIDIA?
Founded in January 1993 and headquartered in California, NVIDIA is an artificial intelligence computing company that designs display chips and is the inventor of GPUs.
NVIDIA is committed to providing first-class chip and software services for gaming, entertainment, PC, cloud data center, networking, and artificial intelligence.
How strong is NVIDIA?
The most intuitive market value comparison: NVIDIA's market value once reached 1 trillion US dollars (about 7.08 trillion yuan), which is equivalent to the A-share Kweichow Moutai (2.1 trillion), + Industrial and Commercial Bank (1.7 trillion) + China Mobile (2 trillion) + CATL (nearly 1 trillion).
It is nearly 2 times the market value of TSMC, 7 times the market value of Intel, 8 times the market value of Qualcomm, and 16 times the market value of SMIC. It is also the world's "most expensive" chip company.
In terms of market share: NVIDIA occupies 88% of the discrete graphics card market, 90% of the AI data center GPU field, and NVIDIA occupies 80% of the global AI market segment.
It can be said that NVIDIA comprehensively crushes Intel and AMD in the field of game graphics cards, GPUs, and AI.
Application scenarios: NVIDIA's GPUs have rapidly penetrated from a single game graphics card to artificial intelligence, supercomputing, quantum acceleration, autonomous driving, the Internet, lithography and other fields.
In the future, with the rapid development of science and technology, NVIDIA will be involved in more and more fields, and even GPU ≥ CPU will appear.
The strongest thing about NVIDIA is the blessing of computing power, and when it encounters the wind of artificial intelligence, it will fly higher and higher.
With the rapid development of the digital economy, future computing power is the cornerstone of the economy, and if you want to develop the economy, you must consolidate the foundation of computing power.
According to IDC research reports, for every 1 point increase in computing power, the country's digital economy and GDP will increase by 3.5% and 1.8%.
When computing power affects GDP, not to mention economic experts, chip giants, even ordinary people are aware of the importance of artificial intelligence chips.
However, a large part of the right to speak of these chips is in the hands of NVIDIA.
Public information shows that 70% of the world's top 500 supercomputers use NVIDIA's GPUs, and the latest supercomputing use rate has reached 90%. And the AI big model is inseparable from NVIDIA.
Therefore, Huang Jenxun declared: in the past decade, the performance of Moore's Law has increased by 100 times, while NVIDIA's GPU performance has increased by 1 million times, and Moore's Law will be invalid in the next ten years, but his own "Huang's Law" will not fail.
Lao Huang directly compared himself to Gordon. Moore even declared that the future will be the world of "Huang's Law", and the status of CPUs in the semiconductor field should also be given to GPUs.
Lao Huang has begun to float, and if it continues like this, Lao Huang will challenge "Newton's laws" one day.
According to Huang Jenxun's algorithm: in the next ten years, NVIDIA can improve the performance of artificial intelligence by 1 million times, so that artificial intelligence can truly awaken.
In order to allow NVIDIA to expand its advantages faster, Huang also launched a new business - "computing power leasing".
We rented houses, rented cars, rented electricity stores, rented girlfriends, and now we can rent "computing power".
For example, you are the owner of a self-media company, with more than a dozen employees, writing articles, shooting videos, and editing videos every day, but the output is limited. What if you want to be more efficient but don't want to hire more copywriters?
Created using AI large models, but general-purpose large models, including ChatGPT, cannot produce high-quality articles and videos.
This requires you to train a specific large model yourself, which is more suitable for specific fields, such as e-commerce, history, tourism, etc.
Training large models requires huge computing power, and the cost of GPT-3 training is about 1.4 million US dollars, which ordinary enterprises simply cannot afford. Rent NVIDIA's computing power to train your own AI model.
$37,000 per month, containing 8 A100 chips, can train ChatGPT in the cloud. Of course, you can also pay a higher price and experience faster training services.
This "computing power leasing" method allows enterprises to reduce the time spent waiting for computing power resources, intervene in artificial intelligence as soon as possible, and carry out related businesses.
In addition, NVIDIA also provides services in the chip field
NVIDIA has released a breakthrough computational lithography technology made in 2nm, the NVIDIA cuLitho computational lithography library.
Chips require a lot of computing in the design and manufacturing process, and as the chip manufacturing process becomes smaller and the number of transistors increases, traditional CPU computing methods become more and more time-consuming.
The use of NVIDIA's GPU solution can increase the computing speed by 40 times, greatly improving the speed of chip design and manufacturing.
For example, a mask that takes two weeks to manufacture can be shortened to 8 hours.
This solution has attracted manufacturing leaders TSMC, EDA leader Synopsys, lithography machine leader ASML and so on.
In the words of Jensen Huang: "AI's iPhone moment has arrived!" Because of NVIDIA. ”
NVIDIA continues to serve China
In August 2022, NVIDIA received a notice restricting the supply of top-of-the-line computing chips for artificial intelligence and data centers to China. Also receiving the notification was AMD Ho Intel.
It is reported that NVIDIA's limited products include A100 and H100 high-performance GPUs.
The A100 is specially tailored for unmanned driving, and NIO, Xpeng, Faraday FF 91, etc. have all used this GPU.
The A100 uses TSMC's 7nm process and integrates 54 billion transistors, and its performance is also 20 times higher than the previous generation V100. And it is a 3D stacked chip with an area of up to 826mm^2 and a maximum power of 400W for the GPU.
In terms of computing power, the A100's 16-bit floating-point computing power reaches 312T, and the 8-bit fixed-point computing power reaches 624T, which can double the computing power in sparse mode.
At the same time, the A100 also supports multi-instance technology, which can divide an A100 into 7 independent small GPUs, so as to realize multiple tasks and calculate at the same time, greatly improving utilization.
H100 is NVIDIA's killer product.
H100 adopts Hopper architecture, TSMC 4nm manufacturing process, integrates 80 billion transistors, has 18432 CUDA cores, 576 Tensor cores, 60MB L2 cache.
In terms of computing power, the H100 6-bit floating-point computing power reaches 1000T, and the 8-bit fixed-point computing power reaches 2000T, which can double the computing power in sparse mode. Compared with the previous generation A100, the computing power has been increased by 3.2 times, and the overall performance has been increased by 6 times.
In terms of data throughput, the H100 can achieve 3TB/s memory bandwidth and 5TB/s Internet speed.
The H100 can also split the GPU, splitting a unit into 7 units for different computing tasks at the same time, and can improve the performance of a single unit by 7 times.
The price of the H100 reached 240,000 yuan, which is worth a good car.
According to the policy of the United States, it is difficult for us to buy A100, H100, what should we do? Nvidia came up with a way to design a GPU specifically for China that meets policy requirements while continuing to occupy the market.
So the A800 and H800 were born, which were used to replace the A100 and H100.
The A800 is still quite conscientious, and the number of cores, single precision, double precision, memory bandwidth, and overall computing power are all completely retained. It's just that the multi-card interconnect performance is affected.
After all, the A100 is a product of 3 years ago, and it is obviously behind the H100, and if it is weakened too much, I am afraid that shipments will be seriously affected.
Compared with the A800, the H800 is a bit excessive, the H800 mainly reduces the data transmission rate to about half of the H100 rate, which is really disgusting to make money and limit others.
In this AI gold rush, Nvidia plays the role of "shovel provider", to the extent that it does not care who can dig up more gold, but how its shovels sell.
If it is sold to American enterprises "shovels" and Chinese enterprises "wooden shovels", the difference is too big, but it will stimulate Chinese companies to vigorously engage in independent research and development, and at the same time, it will be exploited by other shovel companies.
But the United States is different, it wants to ensure that it digs the largest and largest gold mines, so it must limit Chinese companies from getting advanced "shovels", so the future H800 may still reduce computing power.
Nvidia wants to continue to provide computing power services to Chinese companies, while the US government wants to restrict the development of Chinese companies in the field of AI. This policy has led us to wield expensive "broken shovels" to dig broken stones without gold content.
How to solve this situation? Continue to engage in independent research and development!
What is the difference between self-developed AI chips?
In fact, we also have domestic AI chip companies, such as Bicheng Technology, Moore Thread, Jingjiawei, Cambrian, Fangyuan Technology, Hanbo Semiconductor, etc. However, the overall gap with NVIDIA is too big.
Fortunately, one family broke through in terms of computing power, but it was limited in ecology and manufacturing.
At the 2022 World Artificial Intelligence Conference, Shanghai Bicheng Technology released the BR100 series AI chips, which is definitely the pride of domestic chips.
The BR100 uses TSMC's 7nm process and integrates 77 billion transistors. The 16-bit floating point computing power reaches more than 1000T, the 8-bit fixed-point computing power reaches more than 2000T, and the peak computing power of a single chip reaches the PFLOPS level.
BR100 is comparable to NVIDIA H100 in terms of computing power.
However, BR100 has two biggest drawbacks: first, ecological backwardness; Second, manufacturing is limited.
Chip design from difficult to easy: mobile phone SoC> CPU > GPU.
The Kirin series mobile phone SoC designed by Huawei can already be comparable to Qualcomm Snapdragon, and even surpasses Qualcomm in some aspects; Loongson has long developed its own LoongArch architecture, and the gap between performance and Intel has narrowed to 4 years.
So a relatively simple GPU design can't come out? I'm afraid it doesn't make sense!
The domestic GPU lags behind NVIDIA more in terms of ecology, and the GPU + CUDA ecological model established by NVIDIA has become the key to leading.
The CUDA platform is easy to program and powerful, while having a broad and rich ecosystem.
The CUDA platform has available tools, libraries, applications, and partners.
Tools: CUDA toolkit, which provides a development environment using the C/C++ programming language, and a PGI toolkit for programming in Fortran.
Libraries: Libraries can optimize computing architecture, improve software quality, save program development time, and accelerate GPUs.
NVIDIA offers a library called CUDA-X that supports accelerated decline across domains such as linear algebra, image and video processing, deep learning, and graph analysis.
Many partners have also contributed many libraries to the NVIDIA platform.
NVIDIA has many partners, in addition to chip giants such as ARM, TSMC, Intel, Samsung, and Synopsys, as well as Internet giants such as Amazon, FaceBook, Baidu, Alibaba, and Tencent.
Now on the A-shares, once you climb NVIDIA, the stock price will soar.
NVIDIA is too powerful, and domestic AI chip companies are stuck in the ecology, resulting in slow development.
Manufacturing:
Excellent chips are basically manufactured by TSMC, Apple A series, Qualcomm Snapdragon, Huawei Kirin , NVIDIA A, H.
Therefore, even if you design a powerful AI chip, you still have to get TSMC's foundry qualification.
As we all know, after Huawei was included in the entity list, it lost TSMC's foundry qualification. If a Chinese company reaches the level of NVIDIA and affects the interests of Laomei, will it still get the OEM qualification?
The level of SMIC in the mainland is lagging behind TSMC's 3rd generation, which is still under the premise of using imported equipment, which simply cannot meet the foundry needs of advanced AI chips.
Therefore, mainland chip companies are not unable to design advanced AI chips, but they cannot break through ecology and manufacturing.
Write to the end
This month, maybe tomorrow, or next week, Huang will visit China and talk a lot about how to serve Chinese companies and how to ensure the supply of China's advanced AI chips.
At the same time, it will tackle key problems with mainland enterprises and governments to ensure their interests in the Chinese market.
However, the supply guaranteed by NVIDIA is based on the premise of complying with the will of the United States, under the premise of ensuring that China's AI is backward, and under the premise of ensuring that domestic AI chips cannot be surpassed.
If you want to continue to be "stuck in the neck" by others, please accept Huang's advice! Don't forget that on May 29, Huang also said: "There are many GPU (graphics chip) startups in China, don't underestimate China's ability to catch up in the chip field." ”
I am Science and Technology Mingcheng, welcome to discuss together!