Part #/ Keyword
All Products

Samsung Struggles with 3nm AI Chip Yield, Currently at 20%

2024-09-14 14:54:59Mr.Ming
twitter photos
twitter photos
twitter photos
Samsung Struggles with 3nm AI Chip Yield, Currently at 20%

Samsung Electronics is under pressure to improve the yield rates of its advanced AI chips. NVIDIA has indicated that due to strong demand for its chips, it may consider shifting orders away from TSMC.

NVIDIA CEO Jensen Huang recently praised TSMC's "incredible agility and responsiveness to our needs," but also mentioned that "if necessary, we can certainly source from other suppliers." Huang's concerns about AI chip demand are well-founded. He noted that the demand for NVIDIA's Blackwell chips is "enormous," with NVIDIA's suppliers working diligently to meet this demand. However, Huang warned that geopolitical tensions and other concerns might lead NVIDIA to switch suppliers, which could potentially "result in a decline in product quality."

NVIDIA currently outsources the physical production of most of its advanced AI chips to TSMC, rather than other foundries like Samsung. In 2023, Samsung's foundry secured NVIDIA's orders for 8nm chips, such as those used in automotive applications and the RTX-3000 graphics processing units (GPUs) from 2020. However, Samsung has struggled to win orders for more advanced AI chips like the H-series or Blackwell, which require more sophisticated manufacturing processes.

As Huang pointed out, the main issue is quality. Samsung's foundry has been working hard to improve its manufacturing yield, but progress has been slower than expected. The yield for Samsung's 3nm process remained in the single digits until the first quarter of this year, resulting in delays for its own Exynos 2500 chip engineering samples. Analysts believe that while Samsung increased its 3nm yield to nearly 20% in the second quarter, this still falls short of the 60% yield required for mass production.

Amid these challenges, Samsung appears to be struggling to secure foundry orders. Sources suggest that the company has recently revised its equipment installation plans for the P4 wafer fab in Pyeongtaek, Gyeonggi Province, shifting its focus towards producing advanced DRAM memory, such as high-bandwidth memory (HBM) chips, rather than foundry services.

The P4 fab was initially planned to start with NAND equipment, followed by foundry services, and finally DRAM products. However, due to sluggish foundry orders, this sequence has been altered. There are rumors that, given the steady demand for HBM and other advanced storage products for AI servers, Samsung may now dedicate P4 specifically to manufacturing storage chips.

Additionally, Samsung's investment in its Texas-based factory in Taylor is facing scrutiny. The company originally planned to begin mass production of 4nm chips at the facility next year but has postponed this to 2026. While the 4nm process yield is stable, Samsung is working hard to secure orders from fabless companies.

There is speculation that Samsung might shift its focus from 4nm to 2nm chips to ensure future orders for more advanced products. However, reports indicate that the company is also facing challenges in improving the manufacturing yields for both 2nm and 3nm processes.

A semiconductor industry official commented, "One of the core reasons for Samsung's cumulative losses in its foundry business is the low yield rates. While the company has managed to stabilize the 4nm process, it has yet to achieve the same success with more advanced processes such as the second-generation 3nm or 2nm."

Samsung has not disclosed separate earnings for its foundry business, but market observers estimate the company's operating losses for the first half of this year to be approximately 1.5 trillion Korean won (around $1.12 billion).

* Solemnly declare: The copyright of this article belongs to the original author. The reprinted article is only for the purpose of disseminating more information. If the author's information is marked incorrectly, please contact us to modify or delete it as soon as possible. Thank you for your attention!