Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/techinsights-teardown-huawei-ascend-910c-still-contains-cpu-dies-from-tsmc-from-2020.23737/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2030770
            [XFI] => 1060170
        )

    [wordpress] => /var/www/html
)

TechInsights Teardown: Huawei Ascend 910c Still Contains CPU Dies from TSMC from 2020

KevinK

Well-known member

Huawei allegedly smuggled 2.9 million TSMC dies through an intermediary​

Huawei Technologies Co.’s Ascend 910C processors, considered China’s most competitive alternative to Nvidia’s AI hardware, contain advanced components from major Asian chipmakers.

Research firms, including TechInsights conducted teardowns that revealed the Shenzhen-based company relied on advanced parts from Taiwan Semiconductor Manufacturing Co. (TSMC), Samsung Electronics Co., and SK Hynix Inc. to build its third-generation Ascend accelerators.

Huawei relies on stockpiled memory chips for its Ascend 910C​

Huawei began shipping the Ascend 910C earlier this year, giving the market a competitive alternative to Nvidia Corp’s advanced AI processors. While the chips are fully designed in China, investigations conducted by Techinsight revealed that the dies powering the Ascend 910C originated from TSMC’s 7-nanometer process technology, contradicting the assumption that domestic foundry Semiconductor Manufacturing International Corp. had been responsible for fabricating the latest generation.

The South China Morning Post confirmed that separate chip samples contained TSMC-made dies and high-bandwidth memory (HBM2E) obtained from Samsung and SK Hynix. The report confirmed that Huawei gained access to millions of TSMC wafers through an intermediary company called Sophgo.

TSMC has since cut ties with Sophgo, disclosed the transactions to U.S. authorities, and reiterated that it has not supplied Huawei directly since September 2020, according to a Reuters report. However, the existing stockpile of dies is expected to support Huawei’s shipments of the Ascend 910C this year.

Both Samsung and SK Hynix said they ceased doing business with Huawei after export restrictions were imposed. South Korean companies emphasized compliance with U.S. export regulations and confirmed that they no longer supply Huawei with restricted components.

Dylan Patel, a SemiAnalysis analyst, revealed that Huawei purchased approximately $500 million worth of wafers via Sophgo, which later resold around 2.9 million dies to the company. Patel’s analysis suggests that China may face a supply shortage for high-bandwidth memory by the end of this year, with local producers such as Changxin Memory Technologies still far from achieving mass-scale output. SK Hynix is the top producer among Micron Technology and Samsung in developing advanced components used in AI chips.

Ascend 910C combines two Ascend 910B dies​

The Acend 910C processor incorporates HBM2E, an earlier generation of high-bandwidth memory, which is essential to the functioning of AI accelerators. The technology behind HBM is so complex that even Samsung struggled to get its HBM used by Nvidia for years. The 910C processor is the latest generation of Huawei’s earlier series, which combines two 910B dies. The design approach shows Huawei’s reliance on previously acquired foreign technology to extend the life of its hardware lineup.

According to TechInsight research, Ascend 910C has been built on Huawei’s in-house DaVinci architecture. The chip integrates 32 cores capable of delivering 256 teraflops of FP16 performance or 512 TOPS of INT8 performance, supported by 84 MB of on-chip SRAM and four HBM2 channels, which deliver over 1.2 TB/s of memory bandwidth. Unlike most AI accelerators, the 910C also features 16 Arm-compatible Taishan CPU cores, enabling it to operate without a host processor and run complete operating systems.

TSMC has confirmed that the dies used in the recently analyzed chips match those manufactured before the company stopped its shipments, rather than newly fabricated products. Components produced by Samsung and SK Hynix were also identified in separate samples of the 910C. The memory chips, introduced several years ago, are believed to have been acquired before Washington expanded restrictions in 2024 to cover advanced memory sales to China.

The U.S. first added Huawei to its Entity List in 2019, which prevented the company from sourcing advanced semiconductors, manufacturing tools, and design software. Restrictions were expanded in late 2024 to include AI accelerators and the high-bandwidth memory modules that power them. The measures aimed to curb Beijing’s access to frontier AI systems and slow the development of domestic chipmaking capacity.


I don’t have access to the original report, but plenty of commentary on it in the news. Huawei is also living off of banked HBM memory they acquired in 2020.
 
The whole Sophgo claim sounds weird considering they and Huawei are competitors in the AI space. Although Sophgo is more of an edge AI company.

There have been a lot of relatively large die processors announced coming out from China by companies in the Entity List like Loongson.
 
Last edited:
A TechInsights teardown of Huawei's Ascend 910C AI chip found it contains CPU dies from TSMC, manufactured before TSMC stopped supplying Huawei in 2020, confirming the chip utilizes older, stockpiled components rather than the most advanced new technology. This strategy allows Huawei to continue producing its high-end AI chips despite U.S. sanctions, though its long-term production capacity hinges on the remaining stockpile of these foreign-made dies and memory.

Key Findings
  • Older TSMC Dies:
    The Ascend 910C's CPU dies were manufactured by TSMC in October 2024, but they are not newer or more advanced, as TSMC has not supplied Huawei since September 2020.

  • Stockpile Strategy:
    Huawei has been able to produce the 910C by packaging two older TSMC-made dies together, utilizing a significant stockpile of foreign-made components to sustain production despite U.S. export restrictions.

  • Foreign Memory Components:
    The chips also incorporate high-bandwidth memory (HBM2E) from Samsung and SK Hynix, which were also acquired from a stockpile before broader U.S. restrictions on advanced memory shipments to China took effect.

  • Production Limitations:
    The analysis indicates that the availability of these stockpiled components is finite, with the TSMC die stockpile expected to be depleted within about nine months.
Implications
  • Resilience vs. Bottleneck:
    While the use of stockpiled components shows Huawei's resilience in maintaining near-term competitiveness against firms like Nvidia, it also highlights an impending production bottleneck once the stockpiles are exhausted.

  • Supply Chain Dependency:
    The findings underscore China's continued reliance on foreign technology for advanced AI chips and the challenges in rapidly developing domestic alternatives to fill the gap.

  • Regulatory Impact:
    The TechInsights analysis provides further evidence of the complexity and porousness of the U.S. export control measures, as Huawei continues to leverage older technology obtained through intermediaries.
 
There have been a lot of relatively large die processors announced coming out from China by companies in the Entity List like Loongson.
Take a look at the SemiAnalysis report together with TechInsights.

“Currently, all high volume chip production is outsourced to SMIC, the leading Chinese pure-play foundry. This includes the Ascend series of accelerator chips along with the Kirin mobile processors. Yields are poor for SMIC’s 7 nm-class processes due to a combination of immaturity, export controls, and the inherent difficulty in yielding large die such as the Ascend. Thus a relatively low percentage of SMIC’s overall capacity is allocated to producing Ascend die, since smaller mobile processors just make better business sense at this point. But this can change quickly. Let’s discuss what is possible for Huawei….”

SemiAnalysis makes the case that SMIC, and possibly Huawei, can ramp up big die production at 7nm over the next few years to become self-sufficient (I’m a bit more skeptical), but that Huawei will eventually hit the wall in acquiring HBM for AI.


Also thought it was interesting in the video here, that hyperscalers like Alibaba are using key components of NVIDIAs open source AI data center stack for optimizing their data centers.
 
CXMT is working on HBM 3. Allegedly they already sampled it. I doubt that will be an issue.

SMIC doubled its FinFET production area recently so I doubt they won't have the capacity for Ascend. As for SemiAnalysis and Dylan Patel's ideas of what SMIC's 7nm yields are, those are worthless. Dylan has no sources in China.

I also dislike Dylan plastering his logo on images he definitively does not own the copyright for like the satellite foto of CXMT's fab at Hefei or the Kunpeng chip floor plan.
 
Last edited:
CXMT is working on HBM 3. Allegedly they already sampled it. I doubt that will be an issue.

SMIC doubled its FinFET production area recently so I doubt they won't have the capacity for Ascend. As for SemiAnalysis and Dylan Patel's ideas of what SMIC's 7nm yields are, those are worthless. Dylan has no sources in China.
Dylan’s sources are far better than you might imagine. He’s predicted a few things related to China far better than the Digitimes or SCMP.
 
SCMP is a trash Hong Kong publication with mostly British tabloid journalists.
Digitimes from Taiwan is way better. And they are actually trade specific. I would say Nikkei Asia from Japan is also ok.

Of course the best sources are actual Chinese corporate quarterly press releases but not all Chinese companies are in the stock market.

As for Dylan he has lousy sources on anything inside China. And his analysis and speculations are flimsy at best. Often based on false assumptions so GIGO.
I will just give you a couple hints. Huawei's HiSilicon is not the only customer of SMIC's FinFET fab. And that fab complex was designed for 70k WPM.
Unlike his claims the Chinese aren't that interested in the nerfed H20. And unlike his claims you don't need HBM for inference. Even NVIDIA is pushing solutions for that with GDDR.
 
Last edited:
Here is another example of Dylan's superbly flawed analysis:
Huawei currently makes many mobile chips, but there is zero reason to do so from a strategic geopolitical perspective. Oppo and Xiaomi currently fabricate mobile SoCs at TSMC. They are ramping up their own designs to decrease dependence on companies like MediaTek and Qualcomm.
Why the hell would Huawei cede the smartphone market to OPPO or Xiaomi? That is something he never considers.
He always assumes the Chinese semi sector is some kind of central government run top down monolith. Could not be further from the truth.

He also over estimates the current importance of AI for Huawei. They are still a telecoms company first.

ICT Infrastruture and Consumer dwarf the AI revenue from Cloud Computing or Intelligent Automotive by an order of magnitude.
 
Last edited:
SCMP is a trash Hong Kong publication with mostly British tabloid journalists.
Digitimes from Taiwan is way better. And they are actually trade specific. I would say Nikkei Asia from Japan is also ok.

Of course the best sources are actual Chinese corporate quarterly press releases but not all Chinese companies are in the stock market.

As for Dylan he has lousy sources on anything inside China. And his analysis and speculations are flimsy at best. Often based on false assumptions so GIGO.
I will just give you a couple hints. Huawei's HiSilicon is not the only customer of SMIC's FinFET fab. And that fab complex was designed for 70k WPM.
Unlike his claims the Chinese aren't that interested in the nerfed H20. And unlike his claims you don't need HBM for inference. Even NVIDIA is pushing solutions for that with GDDR.
Seriously, you think the best information sources are corporate press releases ? Wouldn't rely on them in any country.
 
Here is another example of Dylan's superbly flawed analysis:

ICT Infrastruture and Consumer dwarf the AI revenue from Cloud Computing or Intelligent Automotive by an order of magnitude.
It's not this simply.

Chinese telecom companies is cutting spending about 10% every year.

The only field with significant sales growth is AI.
 
Even NVIDIA is pushing solutions for that with GDDR.
Only for one piece of transformer-based inference, long-context prefill. CPX is very application-specific, but incredible useful. But it would be advantageous for NVIDIA if Huawei and others try to go down that same path.
 
Unlike his claims the Chinese aren't that interested in the nerfed H20.
There was $5B worth of additional demand, beyond the $12B already sold until the US government interrupted shipments, and then the Chinese government essentially made it illegal for Chinese companies to buy additional units via flimsy claims of “back doors”. Guess what - Chinese companies continue to use the already purchased H20s like nobody’s business. No worries about the the H20s after they were bought.

And it’s not just H20s. US soybeans are also getting nixed by the Chinese government.
 
Last edited:
This sounds like TSMC (also) has a little bit of a 'one eye open, one eye closed' policy with producing chips for China.

Who else but Huawei (through an obvious front) would be using "Huawei CPU dies", that TSMC manufactured in Oct 2024..
 
Because it's the only way to make money.
Sophgo lost so many money.

Sophgo still claims they did not act as a front for Huawei to fab at TSMC. We will find out eventually.
 
Back
Top