Samsung foundry eyes Google TPU component deal

Samsung Electronics may be tapped to produce a key component for Google's next-generation AI chip, the tensor processing unit, according to industry observers. The Taylor fab in Taylor, Texas — where Samsung is in the midst of a major investment push — is expected to serve as the production base.

Google has been pursuing a strategy to diversify its TPU production sources after a surge in orders from major technology companies created a bottleneck at TSMC. Reports that Intel Foundry is also in the running have heightened interest in whether Samsung's foundry division can expand its foothold with Google.

Google is in discussions to have Samsung Electronics' foundry division manufacture a key component of its next-generation TPU, The Information, a US technology publication, reported Thursday.

Google is developing its 10th-generation TPU under the codename Icefish, working alongside Taiwanese chip design firm MediaTek. Mass production is expected to begin as early as 2028. The TPU is Google's proprietary AI chip and has been central to the success of its Gemini AI platform.

Google plans to split production of the 10th-generation TPU between Samsung's foundry division and TSMC, assigning different components to each. The main processor handling computation would use TSMC's 1.4-nanometer process, while the memory input/output die — which connects the processor to high-bandwidth memory — would be manufactured by Samsung's foundry using its 2nm process.

Google reportedly gave high marks to Samsung's deep understanding of memory, given its status as an integrated device manufacturer. The company is believed to have factored in Samsung's comprehensive grasp of HBM and other memory specifications. Samsung is the only company capable of offering a one-stop turnkey solution spanning memory chips, foundry services, system LSI and advanced packaging.

Samsung also supplies the HBM used in Google's TPUs and is locked in fierce competition with SK Hynix for the position of primary supplier. Some in the industry expect Samsung's HBM shipments to Google to surpass those of SK Hynix starting this year.

Google's moves to diversify its TPU production sources have become increasingly visible of late. Reports have emerged that Google, after evaluating Intel's packaging technology, plans to award Intel Foundry production of more than 3 million TPUs starting in 2028 — a volume that would account for roughly half of total output during that period.

TSMC has long held a monopoly on TPU production, but soaring demand for AI chips from Google, Nvidia, Amazon, Microsoft and other major technology companies has created a production bottleneck. Factories running at full capacity are still unable to keep pace with demand.

As Google turns to Intel, Samsung's foundry division is also being mentioned as a potential beneficiary.

Samsung plans to meet demand from US technology companies by scaling up its Taylor fab in Texas. The memory I/O die for Google's 10th-generation TPU is expected to be produced using a 2nm process, and Samsung intends to position the Taylor facility as a hub for that node.

The first fab at the site is set to begin full-scale mass production next year, having completed equipment installation in April and scheduled to begin operation this year. The second fab has entered preliminary planning, proceeding in parallel with discussions to secure orders from global customers. If order volumes grow, investment in the second fab could accelerate.

[단독] “진짜 사장 다시 가려내라” 중노위 재심 봇물…재계 ‘줄소송’ 장기전 비상

https://biz.heraldcorp.com/article/10770198