US Chip Curbs Give Huawei a Chance To Fill the Nvidia Void In China
An anonymous reader quotes a report from Reuters: U.S. measures to limit the export of advanced artificial intelligence (AI) chips to China may create an opening for Huawei to expand in its $7 billion home market as the curbs force Nvidia to retreat, analysts say. While Nvidia has historically been the leading provider of AI chips in China with a market share exceeding 90%, Chinese firms including Huawei have been developing their own versions of Nvidia's best-selling chips, including the A100 and the H100 graphics processing units (GPU). Huawei's Ascend AI chips are comparable to Nvidia's in terms of raw computing power, analysts and some AI firms such as China's iFlyTek say, but they still lag behind in performance. Jiang Yifan, chief market analyst at brokerage Guotai Junan Securities, said another key limiting factor for Chinese firms was the reliance of most projects on Nvidia's chips and software ecosystem, but that could change with the U.S. restrictions. "This U.S. move, in my opinion, is actually giving Huawei's Ascend chips a huge gift," Jiang said in a post on his social media Weibo account. This opportunity, however, comes with several challenges. Many cutting edge AI projects are built with CUDA, a popular programming architecture Nvidia has pioneered, which has in turn given rise to a massive global ecosystem that has become capable of training highly sophisticated AI models such as OpenAI's GPT-4. Huawei own version is called CANN, and analysts say it is much more limited in terms of the AI models it is capable of training, meaning that Huawei's chips are far from a plug-and-play substitute for Nvidia. Woz Ahmed, a former chip design executive turned consultant, said that for Huawei to win Chinese clients from Nvidia, it must replicate the ecosystem Nvidia created, including supporting clients to move their data and models to Huawei's own platform. Intellectual property rights are also a problem, as many U.S. firms already hold key patents for GPUs, Ahmed said. "To get something that's in the ballpark, it is 5 or 10 years," he added.
Read more of this story at Slashdot.