Huawei 910C to Replace Nvidia AI Chips with TSMC and Samsung Tech

Key Takeaways

1. Huawei has allegedly acquired TSMC dies through a shell company, circumventing export controls to produce nearly three million Ascend 910C AI chips.
2. The Ascend 910C chip aims to replace Nvidia AI chips, manufactured at SMIC, but faces challenges in yield rates with the 7nm process.
3. Teardowns confirm the presence of TSMC dies in the 910C, despite TSMC halting production and sales to Huawei post-restrictions.
4. Huawei is set to produce 653,000 Ascend 910C chips but faces difficulties sourcing high-bandwidth memory (HBM), which could hinder future production.
5. Production challenges, including poor packaging and thermal inefficiencies, may limit the output of the 910C to about one million units by 2026, despite significant government funding for local chip production.


Huawei has allegedly found a way to acquire TSMC dies despite existing export control measures, enabling the company to produce its self-developed AI accelerators, as revealed by a detailed examination of the 910C chip.

Ascend 910C: Huawei’s AI Strategy

The Ascend 910C represents China’s primary effort to substitute Nvidia AI chips. Huawei is manufacturing this chip at SMIC while rapidly expanding its own foundry capabilities for full vertical integration. However, with SMIC facing challenges in achieving satisfactory yields in its 7nm production process, Huawei is said to have sourced TSMC dies through a shell company, successfully dodging export restrictions for nearly three million Ascend chips.

Teardown Revelations

Experts in semiconductor teardowns have validated this information, discovering that the 910C chip indeed contains TSMC dies alongside older high-bandwidth memory (HBM) supplied by Samsung and SK Hynix. TSMC, which faced a $1 billion fine due to the leaked dies, was quick to clarify that the recent teardown of the 910C shows it was produced with the same die “examined by this organization in October 2024, not with a newer die or advanced tech.” They also confirmed that the production and sale of chips sent to Huawei were halted after the restrictions came into play.

Production Outlook

Currently, Huawei is on track to produce 653,000 Ascend 910C chips using two TSMC dies, and it reportedly has enough TSMC dies available for production until next summer. However, acquiring high-bandwidth memory to go along with these TSMC chips has proven to be far more challenging. The Samsung or SK Hynix memory that China managed to gather before the HBM controls were implemented is outdated and dwindling fast, leading to reports of Huawei bypassing restrictions by removing memory from loosely packaged products specifically for this task.

Nonetheless, HBM is expected to be the major obstacle for Huawei’s Ascend 910C AI chip production in the future. The 910C delivers about half the performance of the widely-used Nvidia H100 AI chip, which can be found on Amazon for a price comparable to what Huawei reportedly spends to produce the 910C.

Production Challenges Ahead

The 910C’s packaging is somewhat poor and susceptible to thermal inefficiencies, and Huawei still needs to manufacture millions of AI chips to meet domestic demand. Due to these bottlenecks involving HBM and dies, it is estimated that only about a million 910C units could be produced by 2026. However, Chinese government loans aimed at enhancing local AI chip and HBM foundries have reached billions, suggesting that production could ramp up swiftly in the near future.

Source:
Link


 

Comments

3 responses to “Huawei 910C to Replace Nvidia AI Chips with TSMC and Samsung Tech”

  1. เช่ารถรายเดือน avatar

    This post is priceless. When can I find out more?

  2. Hairstyles avatar

    Spot on with this write-up, I really believe this site needs much more attention. I’ll probably be returning to read through more, thanks for the info!

  3. Medium Hairstyles avatar

    It’s great that you are getting ideas from this article as well as from our argument made at this time.

Leave a Reply

Your email address will not be published. Required fields are marked *