With the rapid development of artificial intelligence, AI servers and AI chips have become a sweet spot, and recently, China Mobile announced the winning candidates for the procurement of new intelligent computing centers from 2024 to 2025.
Seven companies, including Kunlun Technology, Huakun Zhenyu, Powerleader Computer, Baixin, Yangtze River Computing, Shenzhou Kuntai, and Xiangjiang Kunpeng, won the bid.
The total amount of winning bids was as high as 19.104 billion yuan, which greatly exceeded market expectations, and refreshed the bidding rate of domestic servers, reaching 100%, which was the first time in the bidding of the three major operators.
It can be expected that in the future, China Telecom, China Unicom, and China Radio and Television will also follow China Mobile's example in the bidding and give priority to the purchase of domestic AI servers.
Many netizens were a little blindsided after seeing the winning bidder, because in the past bidding, Huawei, Inspur, and Shuguang competed, and this time the winning bidder company didn't even hear the name, what's going on?
In fact, there are many domestic server manufacturers, such as Huawei, Inspur, Dawning, Xinhua III, China Great Wall, etc., but their degree of localization is still different.
Huawei's AI server products include the Atlas 500, Atlas 800, and Atlas 900, which are optimized for AI computing and are widely used in cloud computing, scientific research, smart cities, autonomous driving, and other fields.
The server chip uses Huawei's self-developed Ascend AI processor, which ensures security, but the overall performance is lower than that of NVIDIA.
Inspur servers seem to be more capable of playing, but server chips rely heavily on Intel and Nvidia, both of which are chip companies in the United States, which have greatly discounted information security.
Sugon AI servers use processors from Loongson and Haiguang.
We all know that Loongson is a pure domestic chip manufacturer, and has independently developed the LoongArch architecture, which is fundamentally different from X86, ARM, and RSIC-V, and is a pure domestic chip.
However, the Haiguang processor is ostensibly independently developed by Sugon, but it is actually based on the Intel X86 architecture, which has the problems of insufficient security and difficult upgrade and iteration.
H3C servers also rely on chips from Intel, Nvidia and AMD in the United States, and the degree of localization is not high.
The Great Wall of China AI server chip uses Huawei's Ascend, Cambrian, and Bitmain products, and the CPU adopts the Feiteng series.
Cambrian and Bitmain's AI chips are still far behind Huawei, and Feiteng CPU uses ARM V8 architecture, which is not highly autonomous, and whether the future V9 architecture can be authorized is also a problem.
In general, domestic AI servers more or less have the problem of low degree of autonomy and backward performance, and if each company purchases a part of the server, "rain and dew are evenly wet", it may not be able to develop in the end.
It is better to concentrate on supporting the one with the strongest R&D capabilities, the highest degree of autonomy, and the most capable of development, then it is obvious that Huawei is the most suitable.
In the cruel competition, in the future, only Huawei may be able to compete with Nvidia.
Competition between AI chips
Artificial intelligence is known as the fourth industrial revolution, it enters the human world like "water and electricity", and the core base of artificial intelligence is AI chips, so the world's major manufacturers have launched fierce competition.
As a leader in AI chips, Nvidia has a global market share of more than 90%, and AMD has a market share of 4%.
According to IDC data, China's AI accelerator card shipments in 2022 will be about 1.09 million units, with Nvidia accounting for 85%, Huawei accounting for 10%, Baidu accounting for 2%, and Cambrian and Suiyuan Technology both accounting for 1%.
NVIDIA's strong competitiveness benefits from: 1. NVIDIA's computing platform is accelerated, 2. Programmable, and 3. Perfect ecosystem.
NVIDIA is the only company that has continued to cultivate from the original neural network to all subsequent neural networks, deep learning, various vision converters, multimodal converters, and now sequence graph neural networks.
In the entire AI field, time, technology, and talent are the most important, and in the entire AI field, including: games, deep learning, and virtual reality, there are the most customers and the deepest cooperation.
NVIDIA has built a complete ecosystem, including the CUDA ecosystem, intelligent driving platform, small gamified PC ecosystem, data center solutions, etc., and the entire ecosystem has a wealth of development tools, libraries, frameworks, and applications, so that developers can easily develop various AI software.
NVIDIA's technological innovation is also obvious to all, A100, H100, H200, each generation of products can cause the market's frenzied pursuit.
There are also many leading AI chip manufacturers in China, including Huawei, Bichen Technology, Moore Threads, Suiyuan Technology, Muxi, etc., and Bichen Technology also released a BR100 GPU with computing power close to Nvidia H100.
However, under the suppression and restrictions of the United States, the entire domestic AI chip has entered a difficult moment, the BR100 is stuck in the PPT stage, and only Huawei's Ascend series AI chips are still developing and iterating.
Huawei once surpassed Nvidia in the field of AI chips.
As early as August 23, 2019, Huawei officially released the Ascend 910, a commercial AI chip in Shenzhen, when the NVIDIA A100 had not yet been launched, NVIDIA's strongest chip was V100, and the Ascend 910 had about twice the computing power of V100, making it the world's strongest AI chip with the strongest computing power and the fastest training speed at that time.
However, with the inclusion of Huawei in the entity list, Huawei's entire business, including 5G, chips, and smartphones, has been greatly affected, and once reached the embarrassing point where there is no core available, and the strongest 5G equipment manufacturer does not have 5G mobile phones.
In 2023, Huawei Kirin 9000S will come out, which not only adopts a self-developed architecture, but also solves manufacturing problems, marking the return of Huawei chips.
At the same time, Huawei's Ascend 910B is also on the agenda, although there is still a gap with NVIDIA in terms of computing power, but for the domestic market in urgent need of AI computing power, it is tantamount to "a long drought and a rainy day".
The Ascend 910B can benchmark against the NVIDIA A100 in terms of performance.
The Ascend 910B uses Huawei's self-developed Taishan architecture, with a 7nm process, 256 AI cores, and two versions of memory, namely 32GB HBM2 and 64GB HBM2 (high-bandwidth memory Gen 2).
The NVIDIA A100 uses the Ampere architecture, the process is 7nm, has 6912 CUDA cores, the number of transistors has reached 54.2 billion, and there are two versions of memory, namely HBM2 with 40GB and HBM2 with 80GB.
In terms of single-precision (FP32) and half-precision (FP16) computing power, the Ascend 910B is better than the NVIDIA A100.
Overseas large model and cloud computing manufacturers began to purchase NVIDIA H100 and AMD MI300 with 2000TFOLPS high-computing power chips.
Nvidia has also begun to build H200 with 4000T computing power.
It can be said that in the field of AI chips, Nvidia has left the entire Chinese AI chip manufacturers far behind, even if the domestic AI chip industry is struggling to catch up, Nvidia's monopoly position is still unbreakable.
But Huawei, as the biggest dark horse of domestic chips, is running with its head held high, which makes Nvidia also worried, Huang Jenxun bluntly said: In the field of AI, Huawei is one of its main competitors.
The state vigorously supports the domestic chip industry, supports a series of industrial policies, establishes and improves the basic foundation of domestic chips, and accumulates a large number of technical talents.
Talent is the strongest competitiveness of chips, but also the key to China's chips to achieve counterattack and surpass, we have tens of millions of graduates every year, far more than other countries, this is our largest talent treasure trove.
In just a few years, Chinese chips have created a strong R&D team and a complete industrial chain, which is enough to prove that Chinese chips can withstand the test of the market and the cruel competition.
In the future, Chinese chips represented by Huawei will surely enter a new broad road and continue to advance towards "counterattack and transcendence".
I'm a science and technology master, if you like it, give it a like!