China’s race to overcome American chips. How to accumulate the supply chain

Photo of author

By [email protected]


A chip engineer deals with a chip. Dylan Patel, founder, CEO and chief analyst for Semianlysis, compared to the chips restricted at NVIDIA, said the performance gap between Huawei and H20 is less than a whole generation.

Research Moment Gety pictures

With the United States restricting the purchase of advanced semiconductors used in the development of artificial intelligence, Beijing places hopes in local alternatives such as Huawei.

The task has become more challenging with the fact that the United States not only prevents China’s arrival in the world’s most advanced chips, but also restricts biotechnology to create an ecosystem for artificial intelligence.

These restrictions extend to the entire semiconductor chain, from design and manufacturing equipment used to produce artificial intelligence chips to supportive elements such as memory chips.

Beijing may fill Tens of billions One of the dollars trying to fill these gaps, but although it managed to “brute” on its way to some breaches, experts still have their way.

“The US export controls have stimulated the NVIDIA AI chips on China’s industry to develop alternatives,” said Paul Triolo, partner and vice president of China at DGA-Albright Stonebridge.

Here is how China accumulates against the rest of the world in four main sectors required to build artificial intelligence chips.

Artificial intelligence chip design

NVIDIA is the leading artificial intelligence company in the world, but it is important to understand that it does not actually make the physical chips used to train artificial intelligence and computing.

Instead, the company designs artificial intelligence chips, or specifically graphics processing units. Then the patented GPU designs are sent to the company’s chips – manufacturers specializing in mass production of semiconductor products for other companies.

While American competitors such as AMD and Broadcom offer varying alternatives, NVIDIA’s GPU designs are widely recognized as the standard of industry. The demand for NVIDIA chips is so strong that Chinese customers have continued to buy any of the company’s chips that they could get.

But Nafidia is struggling with the restrictions of Washington’s tightening. The company revealed in April Additional restrictions It prevented him from selling the H20 processor to Chinese customers.

The H20 of NVIDIA was a less developed version of its H100 processor, specially designed to avoid previous export controls. However, experts say, he was still more advanced than anything locally available. But China hopes to change this.

In response to restrictions, more Chinese semiconductor players enter the AI. They have included a wide range of events, such as Enflame Technology and Biren Technology, seeking to enjoy Billions of dollars in the GPU graphics processing unit left by NVIDIA.

But no Chinese company appears closer to providing a real alternative to NVIDIA than the Huawei, HISILICON design arm.

The most advanced GPU in Huawei in huge production is 910b. The next generation was 910C It is said Collective shipments are expected to start early in May, although updates have not appeared.

Dylan Patel, founder, CEO and chief analyst at Semianalysis, told CNBC that while the climbs remain behind NVIDIA, they show that Huawei made great progress.

“Compared to the NVIDIA chips for export, the performance gap between Huawei and H20 is less than a full generation. Huawei is not far from the products that NVIDIA is allowed to sell in China,” Patel said.

He added that 910B was behind NVIDIA for a period of two years from last year, while Ascend 910c is only one year later.

But although this indicates that GPU’s design capabilities in China have taken great steps, the design is just one side that stands in how to create an ecosystem for competitive artificial intelligence chips.

Manufacture of artificial intelligence chip

For the manufacture of its graphics processing units, NVIDIA relies on TSMC, the largest chip muse in the world, which produces most advanced chips in the world.

TSMC is compatible with American chips controls, and it is also prevented from taking any chip orders from companies in the American commercial menu. Huawei was placed in the list in 2019.

This has led to Chinese chips designers such as Huawei to recruit local chips, the largest of which is SMIC.

SMIC behind TSMC – It is officially known to be able to produce 7 nm Chips, a technique requires less than TSMC 3 nm produced. Smaller sizes increase the ability and efficiency of chips.

There are signs that SMIC made progress. The company was suspected that the company was behind the 5g Nanomeent 5G chip for Huawei’s Mate 60 Pro, which had it She shook confidence In the United States, the control elements of the chip in 2023. However, the company has a long way before it can produce highly advanced graphics processing units in a cost -cost.

According to the independent analyst for chips and technology, Ray Wang, the well -known SMIC’s operating capacity dwarfed by TSMC.

“Huawei is a very good chips design company, but it is still without good local chips,” Wang said, noting that Huawei It is said Work on its manufacturing capabilities.

But the lack of major manufacturing equipment stands on the path of the two companies.

Advanced chip equipment

SMIC’s ability to meet Huawei’s GPU requirements due to the familiar problem of export controls, but in this case, from the Netherlands.

Although the Netherlands may not have any designer or two semiconductor manufacturers, it is home AsmlThe world’s leading resource for advanced chip industry equipment – machines that use light or electron symptoms to transfer complex patterns to silicon chips, and form the basis of fine chips.

According to US export controls, the country agreed to prevent the sale of the most advanced stone -out of UV printing devices in ASML. Tools are essential for making large -scale advanced graphics processing units.

EUV is the most important barrier to the production of Chinese advanced chips, according to Jeff Koukh, an analyst in Semianlysis. “They have most other available tools, but stone printing limits its expansion capacity and the least practical nodes,” CNBC told CNBC.

I found SMIC Working methods around it Stone printing restrictions using the least dedicated stone print systems in ASML, which have seen relatively lower restrictions.

Through this “brute force”, the production of chips at 7 nm is enforceable, but the returns are not good, and the strategy is likely to reach its maximum, adding that “in current returns, SMIC seems to be not producing enough local accelerators to meet the demand.”

Sicarrier Technologies, a Chinese company that works on collective pilgrimage technology, has It is said He was linked to Huawei.

Koch said that the tradition of current stone printing tools could take years, if not contracts, to achieve it. Instead, China is likely to follow other techniques and various stone printing techniques to push innovation rather than imitation.

Artificial intelligence memory components

While graphics processing units are often determined as the most important components of artificial intelligence, they are far from the only components. In order to run artificial intelligence training and calculation, graphics processing units must work alongside memory chips, which can store data within a broader “set of slides”.

In artificial intelligence applications, a specific type of memory known as HBM has become the standard of industry. South Korea Sk Hynix This industry has been taken in HBM. Other companies in this field include Samsung, based in the United States Micron.

“The memory of the high hesitation range has become at this stage of the progress of artificial intelligence necessary to train and operate models of artificial intelligence,” said analyst Wang.

As with the Netherlands, South Korea cooperated with the US -led chips and began to comply with it New restrictions On selling some HBM memory chips to China in December.

In response to the response, the techniques of Memory Microelectronics Tongfu, in the early stages of HBM production, according to Reuters report.

According to Wang, CXMT is expected to be from three to four years behind global leaders in the development of HBM, although it faces major road barriers, including export controls for chip -making equipment.

Half -solutions The ability in April This remains a year from the activation of any reasonable size.

Wuhan Xinxin Semixonductor Manufacturing It is said Building a factory to produce HBM chips. A a report From SCMP he said that Huawei technologies have partnered with the company in the production of HBM chips, although companies did not confirm the partnership.

HUAWEI has bent on HBM stocks of suppliers like Samsung for use in the ASCEND 910C AI processor. a reportNoting that although the chip is locally designed, it still depends on foreign products obtained before or despite the restrictions.

Half of the solutions said: “Whether it is Samsung HBM or TSMC chips or equipment from America, the Netherlands and Japan, there is a great dependence on the foreign industry.”



https://image.cnbcfm.com/api/v1/image/108157693-1749641770499-gettyimages-1445201644-20221128-a7303615.jpeg?v=1749701249&w=1920&h=1080

Source link

Leave a Comment