China’s efforts to build a domestic AI ecosystem that can challenge US rivals are being ‘hampered by bug-ridden software,‘ a news report says.
China is fostering a domestic AI ecosystem to rival that of dominant US players, but bugs in its alternatives to software from the likes of Nvidia are causing problems.
At issue are tools from Chinese technology giant Huawei, which is developing its own AI chips and also a coding platform, CANN, to rival Nvidia’s CUDA, which speeds the development of applications to take advantage of its GPUs, according to a report in the Financial Times. Software bugs had led to “customers of leading AI chipmaker Huawei complaining about performance issues and the difficulty of switching from Nvidia products,” it said.
[ Related: China’s Moore Threads upgrades AI data center solution amid US restrictions | China seeks 30% growth in national compute capacity by 2025 ]
According to the report, Chinese AI groups are increasingly using Huawei’s Ascend chips to run inference tasks to generate AI responses to queries.
More powerful chips are needed for the initial training of models. However, “multiple industry insiders, including an AI engineer at a partner company, said the chips still lagged far behind Nvidia’s for the initial training of models. They blamed stability issues, slower inter-chip connectivity, and inferior software developed by Huawei.”
Huawei is one of a number fo Chinese companies seeking to meet demand for powerful chips for running AI applications in China, after the US government placed export controls on the technology.
But, Forrester senior analyst Alvin Nguyen said Tuesday, it is unrealistic to expect another organization like Huawei to displace Nvidia quickly, because it will take time and effort from both a hardware and software perspective.
“The difficulties being reported today about Huawei’s software in its attempts to replace Nvidia are unsurprising: The software ecosystem that Nvidia has developed has been around for a long time,” Nguyen said via email. Nvidia introduced CUDA, its software platform for programming GPUs, in 2006.
One researcher told the FT that Huawei’s code made its Ascend chips “difficult and unstable to use,” hampering testing.
“When random errors occur, it is very difficult to find out where it comes from due to poor documentation. You need talented developers to read the source code to see what the issue is, which slows everything down. The coding is imperfect,” the researcher told the FT.
A Chinese engineer with knowledge of Baidu’s use of Huawei’s chips said they crashed frequently, the FT reported.
In its report, the FT said that Huawei has been sending its engineers to customer sites to help them transfer their training code written for Nvidia’s CUDA to Huawei’s CANN.
That could help, said Nguyen: “Huawei using their customer service capabilities to work more closely with customers in using their chips will help them counter Nvidia’s advantage with their software ecosystem long-term.”
The US government is helping too, he said: “Export controls will continue to give opportunities for Huawei to compete, since the export controls prevent Nvidia from selling the ‘best’ version of its products in China.”
Thomas Randall, director of AI market research at Info-Tech Research Group, said, “China is certainly trying to be self-sufficient in advanced chip manufacturing. However, Huawei is suffering yield setbacks in expanding production of its processing hardware, which was originally designed to substitute for Nvidia’s processors.”
Low yields for its processors, he said, “will mean that advanced chip-making will remain commercially unsustainable. This does not mean Huawei is down and out, as they have improved their market share in China — they are just under a continued time crunch to improve their competitiveness.”
And they’re pushing hard. On August 13, The Wall Street Journal reported that Huawei was close to introducing a new chip, the Ascend 910C, and that it was telling potential customers the chip was comparable to Nvidia’s H100, unavailable in China.
