Huawei’s Pursuit of Advanced AI Chips: A Complex Landscape
Despite legal restrictions preventing Huawei from directly acquiring advanced chips from Taiwan Semiconductor Manufacturing Company (TSMC), the company has employed intricate methods to obtain key components for its Ascend 910 AI chips. Recent investigations by TechInsights and TSMC have revealed that Huawei utilized shell companies to procure compute chiplets last year. TSMC, upon discovering the scheme, ceased shipments to Huawei’s intermediaries and initiated an internal inquiry, although the exact quantity of chiplets supplied remains undisclosed.
According to a report by the Center for Strategic and International Studies (CSIS), Huawei managed to secure a significant number of Ascend 910 AI chiplets. The report claims that, “TSMC manufactured large quantities of Huawei Ascend 910B chips on behalf of Huawei shell companies and shipped the chips to China in violation of U.S. export controls.”
Government officials told CSIS that TSMC manufactured more than 2 million Ascend 910B logic dies and that all of these are now with Huawei. If true, this is enough dies to make 1 million Ascend 910C units.
While the report suggests Huawei may possess over two million Ascend 910B logic dies manufactured by TSMC, questions arise concerning the availability of High Bandwidth Memory (HBM) required for integration. The report notes that a potential U.S. plan to restrict advanced HBM sales to China, which was leaked in August 2024 and went into effect in December of that year, provided Huawei with a window to stockpile HBM chips.
It’s important to note, however, that the report does contain some inaccuracies that influence the conclusions reached.
Evolution of the Ascend 910
Huawei’s initial HiSilicon Ascend 910, launched in 2019, comprised a Virtuvian AI chiplet, an Nimbus V3 I/O die, along with four HBM2E memory stacks, and two dummy dies. TSMC manufactured the Virtuvian chiplets for Huawei between 2019 and September 2020, utilizing its N7+ process technology, a 7nm-class node incorporating some extreme ultraviolet (EUV) lithography layers.
Following the U.S. government’s placement of Huawei on its Entity List in 2020, Huawei had to redesign the Virtuvian chiplet. Subsequently, Semiconductor Manufacturing International Corporation (SMIC) produced this revised chiplet using its N+1 technology, which relates to their 7nm-class process. GPUs featuring the new Virtuvian chiplet are designated as HiSilicon Ascend 910B and have no connection with TSMC.
Later, Huawei developed an advanced version of the Virtuvian chiplet for the Ascend 910C. SMIC manufactures the Ascend 910C chiplet using its second-generation 7nm fabrication technology, known as N+2. Contrary to the CSIS report, the Ascend 910C incorporates only a single compute chiplet. Here again, the Ascend 910C is not related to TSMC.
However, Huawei did manage to engage TSMC to produce the original Ascend 910 chiplets for the company in 2023 and into 2024, as TechInsights discovered.
Concerns Regarding Yields
Another element to consider is the somewhat subpar yields of Huawei’s Ascend 910B and Ascend 910C chips. As a result, many of these chips are shipped with certain compute elements disabled. Furthermore, only approximately 75% of Huawei’s AI chips successfully pass advanced packaging procedures.
The CSIS report states:
The advanced packaging process by which two Ascend 910B dies and HBM are combined into a unified Ascend 910C chip also introduces defects that can compromise the functionality of the chip.
Industry sources told CSIS that roughly 75% of the Ascend 910Cs currently survive the advanced packaging process.
Despite these challenges, Huawei continues to procure millions of Ascend 910B and 910C chips for its internal AI initiatives, as well as for external clients. As an example, DeepSeek indicates that the Ascend 910C delivers approximately 60% of the performance demonstrated by Nvidia’s H100. While this level of performance may not be adequate for training extensive language models, it is sufficient for inference workloads.
