90年代申花出租车司机夜晚在车内看文汇报90年代申花出租车司机夜晚在车内看文汇报

Huawei’s Software Woes Hinder China’s Bid to Replace Nvidia inAI

SHANGHAI – China’s ambitions to rival the UnitedStates in artificial intelligence (AI) are being hampered by software glitches, according to industry insiders and reports. Huawei, the leading Chinese AI chipmaker, is facingcriticism from clients who say its software is riddled with bugs and makes it difficult to switch from Nvidia products.

The issue stems from Washington’s tightening ofexport controls on high-performance silicon chips last October, which spurred Chinese tech giants to develop domestic alternatives to Nvidia, the industry leader. Huawei’s Ascend series has become a popular choice for running inference, the process used by applications likeOpenAI’s ChatGPT to generate responses to queries, among a number of Chinese AI technology groups.

However, multiple industry sources, including an AI engineer at a collaborating company, have reported that these chips still lag behind Nvidia’s offeringsin early model training. They attribute this to stability issues, slower chip-to-chip communication speeds, and problems with Huawei’s software, known as CANN.

Nvidia’s software platform, CUDA, is widely regarded as the company’s secret sauce, as it is easy for developers to use andsignificantly speeds up data processing. Huawei is one of many companies trying to break Nvidia’s AI chip monopoly by developing alternative software.

Even Huawei’s own employees have complained about CANN. An anonymous researcher described it as making Ascend products difficult to use and unstable, hindering testing efforts. When random errorsoccur, it is hard to pinpoint the source of the error due to poor documentation. It requires talented developers to read the source code to find the problem, which slows down the work. The coding is not robust, they said.

Another Chinese engineer familiar with Baidu’s use of Huawei processors said the chips frequentlycrash, complicating AI development work. Huawei researchers attributed the crashes to the difficulty of using the hardware. It is easy to get bad results because people don’t understand the hardware itself very well, they said.

To address these issues, Huawei has reportedly dispatched engineers to assist clients in transitioning training code previously writtenon CUDA to CANN. Tech companies like Baidu, iFlytek, and Tencent have all received engineering teams.

Huawei declined to comment. Baidu, iFlytek, and Tencent did not respond to requests for comment.

Huawei is good at customer service, so of course they will send engineers tobig clients to help them use their chips, said a former Baidu employee.

Huawei can leverage its vast workforce to accelerate the transition. The company claims that over 50% of its 207,000 employees are engaged in research and development, including engineers deployed to install technology for clients.

Huawei’s advantage over Nvidia is that it can work closely with customers, said Tilly Zhang, a technology analyst at Gavekal, a consultancy. Unlike Nvidia, Huawei has a large team of engineers who can help customers solve problems and transition to their hardware.

Huawei has also established an online portal for developersto provide feedback on how to improve the software. According to sources familiar with the matter, Huawei raised the price of its Ascend 910B chips used for training by 20% to 30% after the US tightened export controls in October. Huawei’s customers have also expressed concerns about thetight supply of Ascend chips, likely due to difficulties in manufacturing caused by China’s inability to purchase the most advanced chipmaking equipment from ASML in the Netherlands.

Demand for Huawei’s AI chips remains strong. The company reported a 34% increase in revenue in the first half of the year,although it did not provide sales breakdowns for different businesses. Zhang Ping’an, Huawei’s executive director, said in July at the World Artificial Intelligence Conference in Shanghai that over 50 foundational models have been trained and iterated on Ascend chips. iFlytek said last year that Huawei dispatched a teamof engineers to its headquarters in Hefei, eastern China, for technical integration, and its large language model was entirely trained on Huawei chips.

Despite the strong demand and Huawei’s efforts to address the software issues, the challenges remain significant. The company faces an uphill battle to overcome the technical advantages of Nvidia, whichhas a well-established ecosystem and a strong reputation for reliability. The success of Huawei’s AI ambitions will depend on its ability to improve its software and overcome the logistical hurdles imposed by US sanctions.


read more

Views: 0

发表回复

您的邮箱地址不会被公开。 必填项已用 * 标注