Huawei’s buggy software prevents China from replacing Nvidia in artificial intelligence

China’s attempts to reach the level of the United States in terms of computing power in the field of artificial intelligence are hampered by many factors, including problems with software. For example, users of Huawei computing accelerators have complained of performance problems due to software glitches and errors, as well as difficulties in migrating from Nvidia products, the Financial Times reports.

Huawei has emerged as a leader in the race to create a Chinese alternative to Nvidia accelerators after Washington imposed another round of sanctions targeting high-end hardware last October. Ascend accelerators are increasingly being used by Chinese developers to run generative AI models. But in practice, it turns out that Huawei equipment is still far behind Nvidia accelerators at the model training stage. Chinese accelerators are unstable, they have slower inter-chip connections and low-quality Huawei Cann software.

The Cuda software platform is a strong Nvidia product. It is easy to use and can significantly speed up data processing. Huawei is trying to loosen Nvidia’s grip on AI chips by introducing alternative software of its own. But even Huawei employees are complaining about Cann. It makes Ascend’s accelerators “complex and unstable to operate” and makes testing challenging. The quality of the documentation leaves much to be desired, and when errors occur, it is not always possible to determine their source. Huawei Ascend accelerators are used by Baidu, but they often fail, making AI development difficult.

To overcome this problem, Huawei sends its engineers to its largest customers to help transfer the AI ​​training code from Cuda to Cann – the company’s employees work at Baidu, iFlytek and Tencent. More than 50% of Huawei’s 207,000 employees are engaged in research and development, including engineers who help deploy the company’s technology solutions to customers. And this is where Huawei compares favorably with Nvidia. The Chinese company has even launched an online portal for developers, where they can leave suggestions for improving the software.

After the introduction of another package of American sanctions, Huawei raised prices for Ascend 910B accelerators, which are used for AI training, by 20–30%. At the same time, their supply volumes are limited, probably due to production difficulties – Chinese companies cannot purchase equipment for the production of chips from the Netherlands ASML. Huawei AI accelerators are in high demand. In the first half of the year, the company increased revenue by 34%, but did not provide a breakdown by business area. More than 50 basic AI models have already been trained and integrated on Ascend accelerators, Huawei estimates – iFlytek reported that it trained its neural network exclusively on Huawei equipment, and its manufacturer sent a group of engineers to the company to help deploy solutions.