top of page

China's AI Hurdle

Huawei's ambitious efforts to compete with Nvidia in the AI chip market face significant challenges, as software bugs and stability issues plague its Ascend series chips. While the Ascend 910B shows promise in raw computing power, frequently matching or even surpassing Nvidia's A100 in some tests, persistent problems with Huawei's Cann software platform and manufacturing constraints at SMIC hinder its market competitiveness. These issues not only impact Huawei's partnerships with major tech companies but also raise questions about China's broader ambitions for semiconductor self-sufficiency in the face of U.S. export restrictions.


Huawei's AI Chip Challenges 

The Ascend series chips face significant stability issues, including frequent crashes and slower inter-chip connectivity, hampering their performance in AI model training.These problems stem from bugs in Huawei's Cann software platform, which lacks comprehensive documentation and user-friendliness compared to Nvidia's CUDA.Manufacturing constraints at SMIC further complicate matters, with only about 20% of produced chips functioning as intended due to limitations in the 7nm process technology. These challenges have led to a reduction in the number of active AI cores in the Ascend 910B compared to its predecessor, likely due to poor yields or limited production capacity.


Comparison: Ascend 910B vs. A100 

The Ascend 910B demonstrates competitive performance against Nvidia's A100, achieving up to 80% of its efficiency in large language model training and even surpassing it by 20% in certain scenarios. However, the 910B consumes more power at 400W compared to the A100's 300W, potentially impacting its efficiency in power-sensitive environments. While the 910B shows promise in raw computing power, its overall effectiveness is hampered by software instability and lower yields, which affect reliability and availability in real-world applications.


Huawei's Customer Support Strategies 

To address the stability issues with its AI chips, Huawei has implemented a hands-on customer support strategy that sets it apart from competitors like Nvidia. The company deploys engineers directly to customer sites to assist with the transition from Nvidia's CUDA to Huawei's Cann platform and to troubleshoot issues on-site. This approach leverages Huawei's large R&D workforce, with over half of its 207,000 employees working in research and development. Additionally, Huawei has opened an online portal for continuous customer support and feedback, aiming to improve its software and hardware based on real-world usage and challenges.


Future Plans for AI Chips

Plans are underway to introduce the Ascend 910C, an advanced successor to the 910B, and to develop processors using a more sophisticated 5nm process. These advancements aim to improve performance and yield rates, potentially addressing current manufacturing challenges. However, ongoing equipment and supply chain issues, exacerbated by U.S. export controls, may continue to pose significant obstacles to scaling up production for these next-generation processors. The success of these future chips will be crucial for Huawei's competitiveness in the global AI market and China's broader ambitions for semiconductor self-sufficiency.



If you work within a wine business and need help, then please email our friendly team via admin@aisultana.com .


Try the AiSultana Wine AI consumer application for free, please click the button to chat, see, and hear the wine world like never before.



4 views0 comments

Recent Posts

See All

Use of Facial Recognition Surges

Deployments of live facial recognition technology have led to arrests for serious crimes, including rape and robbery.

Cerebras, an A.I. Chipmaker

Cerebras Systems, a pioneering AI chipmaker, has developed the Wafer Scale Engine 3 (WSE-3), a groundbreaking chip.

LiquidAI Debuts GPT Rival

Liquid Foundation Models (LFMs), a new series of AI models that challenge traditional large language models like GPT.

Comentarios


Los comentarios se han desactivado.
bottom of page