Huawei's Ascend 910C AI Chip Production Yield Improvement and Market Impact

Significant Enhancements in Production and Performance / Reuters

Huawei has reportedly improved the production yield of its Ascend 910C AI chip from 20% to nearly 40%, marking a crucial advancement in its commercial viability. This increase in yield indicates a higher percentage of functional chips produced during the manufacturing process, thereby enhancing cost-effectiveness and market competitiveness. Notably, the Ascend 910C is assessed to achieve approximately 60% of the inference performance of Nvidia's H100 GPU, a finding that highlights Huawei's focus on inference tasks rather than training applications. This development suggests a strategic move towards reducing reliance on Nvidia within China's AI applications.

The Chinese government is actively encouraging domestic firms to adopt Huawei's AI chips, further bolstering the company’s market share. The progress in production and performance signifies a substantial step towards achieving self-sufficiency in AI computing infrastructure, despite the ongoing U.S. export restrictions.

Recent reports indicate that Huawei’s partnership with SMIC (China's leading semiconductor manufacturer) has facilitated this yield improvement. By employing the N+2 manufacturing process, which does not utilize extreme ultraviolet (EUV) technology, Huawei has managed to enhance its chip production capabilities significantly. As a result, analysts from Creative Strategies note that a 40% yield for the Ascend 910C positions it well for commercial deployment, especially when compared to TSMC's 60% yield for the Nvidia H100 AI processor.

In terms of performance, the Ascend 910C's ability to reach 60% of the Nvidia H100’s inference capacity suggests that Huawei has effectively designed the chip for specific AI tasks. While some sources claim that the Ascend 910C could match or even surpass the H100's performance in certain scenarios, comprehensive benchmark data remains limited. Reports indicate that the Ascend 910C delivers 320 TFLOPS in FP16 performance and 64 TFLOPS in INT8 performance, indicating its suitability for various AI applications, although it may still lag behind in overall training capabilities compared to the H100.

For 2025, Huawei plans to produce 100,000 units of the Ascend 910C and 300,000 units of the Ascend 910B, totaling over 400,000 AI chips. Given the current yield rate, achieving these production targets will require a substantial number of wafers. For instance, assuming a die size similar to that of the H100 (approximately 800mm²), around 80 chips can be produced per 300mm wafer. With a 40% yield, this translates to roughly 32 viable chips per wafer, necessitating approximately 3,125 wafers to manufacture the 100,000 units of the Ascend 910C.

Moreover, Huawei commands over 75% of the total AI chip production in China, reinforcing its dominant position in the domestic market. This statistic reflects the increasing adoption of its AI technologies amidst a broader governmental push to substitute imported chips with local alternatives.

In response to U.S. export controls, Huawei is unable to procure cutting-edge EUV lithography equipment from ASML, pushing SMIC to rely on DUV technology for its manufacturing processes. Although this may place Huawei at a technological disadvantage compared to TSMC’s latest advancements, the company’s founder and CEO, Ren Zhengfei, expressed optimism about enhancing China's technological self-sufficiency, noting a reduction in concerns regarding the nation’s technological capabilities in front of President Xi Jinping.

The proactive testing and purchasing negotiations for the Ascend 910C by major firms such as ByteDance, Baidu, and China Mobile underscore Huawei’s substantial market potential. However, the ambitious production plans raise questions about the sustainability of achieving over 400,000 chips, especially considering the mixed reports regarding yield rates. While some sources assert that the yield may still be as low as 20%, others point to improvements that could enhance credibility and transparency regarding Huawei's technological data.

A summary table comparing key chip yields and performance metrics is as follows:

Chip Name	Manufacturer	Yield (%)	Inference Performance (vs. H100)	Production Plans (2025)
Ascend 910C	SMIC	~40	60%	100,000 units
Ascend 910B	SMIC	~50	N/A	300,000 units
Nvidia H100	TSMC	~60	100%	N/A

The enhancements in Huawei's Ascend 910C yield and performance signify a pivotal development for the Chinese AI industry, promoting greater self-reliance in technology. However, the challenges posed by technological gaps with advanced foundries like TSMC and the complexities of scaling production capacity remain significant factors in the evolving landscape of U.S.-China technological competition.