Since the beginning of the Trump administration, the United States has been restricting and tightening exports of chips to China, forcing the Chinese Communist Party to invest heavily in an attempt to develop leading chip capabilities. However, the results have been minimal.
Huawei has been trying to compete with Nvidia in the artificial intelligence (AI) chip market. Recently, Chinese enterprises and industry insiders revealed that Huawei’s AI chips are riddled with flaws, lagging significantly behind Nvidia in terms of technicality and stability.
According to a report by the Financial Times of the UK, multiple industry sources have indicated that Huawei’s Ascend series of AI chips lack stability in connectivity and training efficiency falls far behind Nvidia. Additionally, Huawei’s heterogeneous computing architecture CANN computing platform, targeted at AI scenarios, has been criticized for its defects as an alternative software originally intended to replace Nvidia’s CUDA.
Nvidia has maintained a leading position in the AI chip field, especially in high-performance computing, with its products being the top choice for global data centers and tech companies. The CUDA computing platform is a comprehensive ecosystem that Nvidia has spent years developing in the high-performance computing field, integrating computing platforms, programming models, and development tools.
Last October, Washington further tightened export controls on high-performance chips to China, with hopes pinned on Huawei by the Chinese authorities as a replacement for industry leader Nvidia.
Huawei’s Ascend series is becoming increasingly popular in the AI field in China, used for executing inference processes, similar to the way OpenAI’s ChatGPT generates responses to queries.
However, under the US technology blockade policy, Huawei is unable to access the technology, equipment, and raw materials needed for producing advanced chips, resulting in limitations on its production efficiency and product quality.
Many industry insiders, including an AI engineer from a company that collaborates with Huawei, have expressed that Huawei’s chips significantly lag behind Nvidia in the early stages of model training. They attribute this to stability issues, slower chip-to-chip connection speeds, and Huawei’s poor-quality CANN software.
Even Huawei’s employees have been complaining about CANN. A research scientist who preferred to remain anonymous stated that CANN makes Ascend products “difficult to use and unstable,” hindering testing work.
Another Chinese engineer familiar with Baidu’s use of Huawei software mentioned that these chips often crash, complicating AI development work.