Stay up to this point with free updates
Simply register for Artificial intelligence myFT Digest – delivered straight to your inbox.
China's efforts to match U.S. computing power in artificial intelligence are being hampered by buggy software, with customers of leading AI chipmaker Huawei complaining about performance issues and the problem of switching from Nvidia products.
The Chinese technology giant has emerged because the frontrunner within the race to develop a domestic alternative to industry leader Nvidia after Washington further tightened export controls on high-performance silicon last October.
The Ascend series has change into an increasingly popular option for Chinese AI groups to perform inference, a process that applications like OpenAI's ChatGPT use to generate responses to queries.
But several industry insiders, including an AI engineer at a partner company, said the chips still lagged far behind Nvidia's in initially training models, blaming stability issues, slower inter-chip connectivity and inferior software developed by Huawei called Cann.
Nvidia's Cuda software platform is taken into account the corporate's “secret sauce” since it is simple for developers to make use of and may speed up data processing enormously. Huawei is certainly one of many firms attempting to break Nvidia's dominance in AI chips by developing alternative software.
Huawei's own employees are also complaining about Cann. One researcher, who wished to stay anonymous, said it made the Ascend product “difficult and unstable to make use of” and hindered work on its testing.
“When random errors occur, it's very difficult to determine where they're coming from due to poor documentation. You need talented developers reading the source code to determine the issue, which slows all the things down. The coding isn't perfect,” they said.
Another Chinese engineer briefed on Baidu's use of Huawei processors said the chips crash incessantly, making AI development work tougher.
The Huawei researcher said crashes occurred since the hardware was difficult to make use of. “It's easy to get bad results because people don't know much in regards to the hardware itself,” they said.
To solve the issue, Huawei has sent engineers to assist customers on site transfer training code previously written on Cuda to Cann, in keeping with several people conversant in the matter. Baidu, iFlytek and Tencent are among the many technology firms which have received engineering teams, these people said.
Huawei declined to comment. Baidu, iFlytek and Tencent didn’t reply to requests for comment.
A former Baidu worker said: “Huawei has an impressive customer support record. So after all they’ve engineers on site at their major customers to assist them deploy their chips.”
Huawei can speed up change with an enormous workforce. According to the corporate, greater than 50 percent of its 207,000 employees work in research and development, including the engineers who’re sent to put in the technology at customers' sites.
“Huawei's advantage over Nvidia is that it may well work closely with its customers,” said technology analyst Tilly Zhang of consulting firm Gavekal. “Unlike Nvidia, Huawei has a big team of engineers who can assist customers solve their problems and enable them to migrate to their hardware.”
Huawei has also arrange an internet portal where developers can provide feedback to enhance the software.
After the U.S. tightened export controls in October, Huawei raised the worth of the Ascend 910B, the chip used for training, by 20 to 30 percent, people conversant in the matter say.
Huawei customers also expressed concerns about supply shortages of the Ascend chip, that are believed to be because of production difficulties as Chinese firms are prevented from purchasing state-of-the-art chip-making machines from Dutch company ASML.
Huawei is seeing strong demand for its AI chips. On Thursday, the corporate reported a 34 percent increase in sales in the primary half of the 12 months, but didn’t provide a breakdown of sales by business unit.
More than 50 basic models have been “trained and iterated” on the Ascend chip, said Huawei CEO Zhang Ping'an on the World Artificial Intelligence Conference in Shanghai in July.
iFlytek said its large language model was trained exclusively on Huawei chips after Huawei sent a gaggle of engineers to its headquarters in Hefei, eastern China, last 12 months to integrate the technology.
