China’s racing to beat U.S. chip curbs. How its supply chain stacks up

The United States restricting China to buy advanced semiconductors used in the development of artificial intelligence, Beijing places hopes on domestic alternatives such as Huawei.

The task was made more difficult by the fact that the American limits not only inhibit China’s access to the most advanced fleas in the world, but also limit vital technology to create a ecosystem of AI fleas.

These constraints extend over the entire semiconductor value chain, ranging from the design and manufacturing equipment used to produce AI chips to support elements such as memory chips.

Beijing mobilized tens of billions Dollars to try to fill these shortcomings, but even if she was able to “force” her way in certain breakthroughs, he still has a long way to go, according to experts.

“US export controls on NVIDIA AI advanced chips have prompted Chinese industry to develop alternatives, while making national companies more difficult to do,” said Paul Triolo, partner and vice-president of China to the consulting company DGA-Albright Stonebridge Group.

Here is how China is accompanied by the rest of the world in four key segments necessary to build AI fleas.

But Nvidia is struggling with Washington’s tightening restrictions. The company revealed in April that Additional borders had prevented her from selling his H20 processor to Chinese customers.

The NVIDIA H20 was a less sophisticated version of its H100 processor, designed specifically to bypass the previous export controls. Nevertheless, say the experts, it was even more advanced than anything available at the national level. But China hopes to change this.

In response to restrictions, more Chinese semiconductor players have entered the AI ​​processor arena. They included a wide vendor of upstarts, such as technology ignited and Biren technology, seeking to soak up Billions of dollars in the request of GPU left by Nvidia.

But no Chinese company appears closer to providing a real alternative to Nvidia than Huawei’s design arm, Histilicon.

Huawei’s most advanced GPU in mass production is its 910B ascent. The New generation 910C Ascend was would have should start mass expeditions from May, although no update has emerged.

Dylan Patel, founder, CEO and chief analyst in semianalysis, told CNBC that even if the ascending chips remain behind Nvidia, they show that Huawei has made significant progress.

“Compared to export fleas at Nvidia export, the performance gap between Huawei and the H20 is less than a complete generation. Huawei is not far behind the products that Nvidia is authorized for sale in China,” said Patel.

He added that the 910B was two years late on Nvidia from last year, while Ascend 910c is only a year late.

But although this suggests that the design capacities of the GPU of China have made great progress, design is only one aspect which is not part of the creation of a competitive chip ecosystem.

SMIC is far behind TSMC – he is officially known to be able to produce 7-Nanomier Chips, requiring less advanced technologies than the production of 3 TSMC nanometers. Smaller nanometers cause greater power to process fleas and efficiency.

There are signs that Smic has made progress. The company is suspected of having been behind a 5g chip of 5 nanometers for Huawei’s Mate 60 Pro, who had confidence In orders for American fleas in 2023. The company, however, has a long way to go before being able to mass produce advanced GPUs in a profitable manner.

According to the independent analyst of chips and Ray Wang technology, the known operating capacity of SMIC is overshadowed by TSMC.

“Huawei is a very good flea design company, but they are still without good manufacturers of national fleas,” said Wang, noting that Huawei is would have work on your own manufacturing capacities.

But the lack of key manufacturing equipment hinders the two companies.

While the Netherlands may not have leading designers or manufacturers of semiconductors, it houses ASMLThe world’s leading supplier of advanced Tamias manufacturing equipment – machines that use light or electronic beams to transfer complex patterns on silicon platelets, forming the base of microchips.

In accordance with US export controls, the country has agreed to block the sale of ASML (EUV) ultraviolet lithography machines (EUV). The tools are essential to make advanced GPUs on a large scale and in a profitable manner.

The EUV is the most important barrier for Chinese production of advanced fleas, according to Jeff Koch, semi-analysis analyst. “They have most of the other tools available, but lithography limits their ability to evolve around 3 nm and less process nodes,” he told CNBC.

SMIC found Methods to bypass Lithography restrictions using less advanced ultraviolet lithography systems of ASML, which have seen relatively fewer restrictions.

Thanks to this “raw forcing”, the production of fleas at 7 Nm is doable, but the yields are not good, and the strategy probably reaches its limit, said Koch, adding that “to current yields, it seems that SMIC cannot produce enough national accelerators to meet demand”.

Sicarrier Technologies, a Chinese company working on lithography technology, has would have been linked to Huawei.

But imitating existing lithography tools could take years, even decades, to realize, said Koch. Instead, China is likely to pursue other technologies and different lithography techniques to push innovation rather than imitation, he added.

In AI applications, a specific type of memory known as HBM has become the standard of industry. In South Korea SK Hynix took the industry in HBM. Other companies in the field include Samsung and the United States Micron.

“The high memory of the bandwidth at this stage of AI progression has become essential for the training and management of AI models,” said analyst Wang.

As with the Netherlands, South Korea cooperates with the restrictions of the flea led by the United States and began to comply with edge When selling certain HBM memory chips in China in December.

In response, the Chinese manufacturer of the memory of fleas Changxin Memory Technologies, or CXMT, in partnership with the Package and Tongfu Microelectronics Package Society, is at the early stages of HBM production, according to a Report by Reuters.

According to Wang, CXMT should have three to four years behind the world leaders in the development of HBM, although it is confronted with major roadblocks, including export controls on Tamias manufacturing equipment.

Semianalysis estimated in April This CXMT remained one year from the slaughter of any reasonable volume.

Foundry Chinese Wuhan Xinxin Semiconductor Fabrication is would have Build a factory to produce HBM plates. A report De SCMP, said Huawei Technologies has teamed up with the company in the production of HBM fleas, although companies have not confirmed the partnership.

Huawei relied on the HBM stocks of suppliers like Samsung for use in their Ascend 910C AI processor, said semiianalysis in April reportNoting that if the chip has been designed at the national level, it is still based on foreign products obtained before or despite the restrictions.

“Whether HBM from Samsung, TSMC wafers or America, the Netherlands and Japan equipment, there is great dependence on the foreign industry,” said Semianalysis.