Analysis: US chip restrictions give Huawei a chance to fill Nvidia’s gap in China

FILE PHOTO: The logo of Huawei Technologies is seen at its exhibition space during the Viva Technology conference dedicated to innovation and startups at the Porte de Versailles exhibition center in Paris, France, June 15, 2022. Reuters/Benoit Tessier/File Photo/File Photo Acquire License Rights

HONG KONG, Oct 20 (Reuters) – U.S. moves to limit exports of advanced artificial intelligence (AI) chips to China may present an opportunity for Huawei technology. [RIC:RIC:HWT.UL] In its $7 billion home market, analysts say, as restrictions force Nvidia to pull back.

While Nvidia ( NVDA.O ) has historically been the leading provider of AI chips in China, with a market share of more than 90 percent, Chinese companies including Huawei are developing their own versions of Nvidia’s best-selling chips, including the A100 and H100 graphics. . processing units (GPU).

Analysts and some AI companies such as China’s iFlyTek ( 002230.SZ ) say Huawei’s Ascend AI chips are comparable to Nvidia’s in terms of raw computing power, but still lag behind in performance.

Another key limiting factor for Chinese companies is the reliance of most projects on Nvidia’s chips and software ecosystem, but that could change with US restrictions, said Jiang Yifan, senior market analyst at brokerage Guotai Junan Securities.

In a post on his Weibo account, Jiang said, “I think this US move is actually giving Huawei’s Ascend chips a big gift.”

However, this opportunity comes with several challenges.

Many advanced AI projects are built with CUDA, a popular programming architecture pioneered by Nvidia, which in turn has spawned a massive global ecosystem capable of training highly complex AI models such as OpenAI’s GPT-4.

Huawei’s own version is called CANN, and analysts say it’s much more limited in terms of the AI ​​models it can train, meaning Huawei’s chips are far from a plug-and-play replacement for Nvidia.

Woz Ahmed, a former chip design director turned consultant, said that in order for Huawei to win Chinese customers from Nvidia, it needs to replicate the ecosystem created by Nvidia, including supporting customers to transfer their data and models to Huawei’s own platform.

Intellectual property rights are also an issue, Ahmad said, as many US companies already hold key patents for GPUs.

“It takes five or 10 years to get something that’s in the ballpark,” he added.

Huawei and Nvidia did not immediately respond to Reuters’ request for comment.

Perks

If Huawei can capture Nvidia’s market share, it could score another victory against the United States, which has targeted export controls since 2019.

Huawei released the first Ascend GPUs that year, and it’s one of a number of products — like its Harmony operating system — that the company says is entirely in-house.

Over the past year, the telecom giant has shown signs of breaking away from U.S. restrictions by unveiling an advanced smartphone chip and claiming advances in chip design tools.

The company is also aiming to become a key supplier of computing power for artificial intelligence, with Chief Financial Officer Meng Wanzhou saying last month that Huawei wants to build a computing base for China, in a confidential reference. Give the world a “second option”. to the dominant US provider

Huawei’s partners in China so far include iFlyTek, a leading Chinese AI software company, which uses the Ascend 910 to train its AI models. IFlyTek was also blacklisted by the US in 2019.

Jiang Tao, iFlytech’s senior vice president, said Thursday that the capabilities of the Ascend 910B are “comparable to Nvidia’s A100,” and announced that it is developing general-purpose AI infrastructure in China alongside Huawei.

“Right now the goal of our partnership is to make it possible to build domestically developed LLMs with domestic hardware and software technology,” Jiang said.

Other partners include state-owned software companies Tsinghua Tongfang and Digital China. At a conference in July, Huawei said its AI chips are now helping more than 30 large language models (LLMs) in China, which is going through an AI production frenzy and currently has more than 130 LLMs.

86Research analyst Charlie Chai says Nvidia’s dominance of the ecosystem “is not an insurmountable obstacle if domestic players are given enough time and a large customer base.”

China’s push for self-sufficiency, championed by President Xi Jinping, is likely to help. “In short, a small disruption to short-term supplies, but a big boost to the long-term self-sufficiency plan,” Chai added.

($1 = $10,000)

Our Standards: The Thomson Reuters Trust Principles.