Huawei’s AI Chips Face Performance Hurdles in Nvidia Challenge

AI chip competition intensifies: Huawei’s efforts to develop a domestic alternative to Nvidia’s AI chips are facing significant challenges, with customers reporting performance issues and difficulties in transitioning from Nvidia products.

• Huawei’s Ascend series has become increasingly popular for running inference in Chinese AI applications, particularly after Washington tightened export controls on high-performance silicon in October.
• However, industry insiders report that Huawei’s chips still lag far behind Nvidia’s for initial model training, citing stability issues, slower inter-chip connectivity, and problems with Huawei’s software platform, Cann.
• Nvidia’s software platform, Cuda, is widely regarded as the industry standard, known for its ease of use and ability to accelerate data processing significantly.

Software challenges: Huawei’s Cann software platform is proving to be a major hurdle in the adoption of its AI chips, with users reporting numerous issues and difficulties in implementation.

• Huawei employees have described Cann as “difficult and unstable to use,” with poor documentation making it challenging to identify and resolve errors.
• Engineers working with Huawei’s processors report frequent crashes, complicating AI development work and slowing down progress.
• The complexity of using Huawei’s hardware has been cited as a contributing factor to these issues, with users struggling to achieve optimal results due to a lack of familiarity with the technology.

Huawei’s response: To address these challenges, Huawei is taking proactive steps to support its customers and improve its technology.

• The company has been dispatching teams of engineers to assist customers in transitioning their training code from Cuda to Cann, working on-site with major tech companies such as Baidu, iFlytek, and Tencent.
• Huawei’s large workforce, with over 50% of its 207,000 employees working in research and development, allows it to provide extensive customer support and accelerate the adoption of its technology.
• The company has also established an online portal for developers to provide feedback on improving its software, demonstrating a commitment to addressing user concerns and enhancing its products.

Market dynamics: Despite the challenges, Huawei is seeing strong demand for its AI chips, driven by the need for domestic alternatives to US-made products.

• Following the tightening of US export controls, Huawei increased the price of its Ascend 910B chip, used for training, by 20 to 30 percent.
• Customers have expressed concerns about supply constraints for the Ascend chip, likely due to manufacturing difficulties stemming from restrictions on purchasing advanced chipmaking machinery.
• Huawei reported a 34% increase in first-half revenues, though specific sales figures for its AI chip business were not disclosed.

Industry adoption: Despite the challenges, Huawei’s AI chips are gaining traction in the Chinese market, with several companies integrating the technology into their operations.

• More than 50 foundational models have been trained and iterated on the Ascend chip, according to Huawei executive director Zhang Ping’an.
• iFlytek has reported that its large language model has been trained exclusively on Huawei chips, following collaboration with Huawei engineers to integrate the technology.

Broader implications: The struggles faced by Huawei in developing competitive AI chips highlight the complexities of the global semiconductor industry and the challenges of creating alternatives to established technologies.

• The situation underscores the significant lead that Nvidia maintains in the AI chip market, particularly in terms of software ecosystem and ease of use.
• It also demonstrates the potential long-term effects of export controls on technological development and competition in the AI sector.
• As Huawei continues to refine its technology and support its customers, the success of its AI chip initiative could have significant implications for the future of AI development in China and the global balance of technological power.