"Complaints Mount" Huawei AI Chip... Can It Become an 'NVIDIA Alternative'?

"Performance Lagging Behind Nvidia and 'Can' Unstable"

Despite China's efforts to strengthen its own capabilities in response to the U.S.'s advanced semiconductor sanctions, it is still insufficient to replace Nvidia's artificial intelligence (AI) chips, major foreign media reported on the 2nd (local time). Huawei's Ascend series is widely used, but complaints from customers are emerging one after another.

After the U.S. tightened export controls on advanced semiconductors to China, Chinese companies, which lost access to Nvidia's advanced chips, have rushed to develop alternatives. In particular, Huawei's Ascend series is considered a strong contender against Nvidia.

[Image source=AFP Yonhap News]

Last month, The Wall Street Journal (WSJ) reported that Huawei plans to launch the Ascend 910C AI chip as early as October, and leading companies such as ByteDance, the parent company of TikTok, search engine Baidu, and state-owned telecom China Mobile have ordered Ascend 910C chips worth $2 billion (approximately 2.6818 trillion KRW). Huawei reportedly told its customers that the Ascend 910C chip is comparable to Nvidia's latest chip, the H100.

However, according to industry insiders including AI engineers from partner companies, Huawei's Ascend series lags behind Nvidia in performance during the initial training of AI models. Industry insiders also raised issues regarding stability, chip interconnectivity, and Huawei's developer tool 'CANN'.

Nvidia's developer tool 'CUDA' is easy to use and can quickly process large-scale data. It is credited as a key factor that allowed Nvidia to dominate the global AI chip market.

Another Chinese engineer who received a briefing from Baidu on the use of Huawei processors revealed that the chips frequently crash, complicating AI development work. Complaints about CANN have also reportedly emerged within Huawei. A researcher who requested anonymity stated that CANN made using Ascend products difficult and unstable, hindering testing tasks.

According to multiple sources, Huawei dispatched engineers to major customers such as Baidu, iFlytek, and Tencent to assist with transferring training codes written in Nvidia CUDA to CANN in order to resolve these issues. Huawei also established an online portal where developers can provide feedback on how to improve its software.

A former Baidu employee said, "Because Huawei has excellent customer service, it naturally deploys engineers on-site at large customers to assist with chip usage." According to Huawei, more than 50% of its 207,000 employees are R&D personnel, including engineers dispatched for technical installations.

Tilly Zhang, a technology analyst at consulting firm Gabcal, said, "Huawei's advantage over Nvidia is that it can work closely with customers," adding, "Unlike Nvidia, Huawei has a large engineering team that solves customer problems and helps customers transition to its hardware."