• 구독신청
  • My스크랩
  • 지면신문
FNTIMES 대한민국 최고 금융 경제지
ad

'AI Core' HBM4 Competition Strategy Poles Apart: Samsung Electronics 'Solo' vs SK Hynix 'Alliance'

곽호룡 기자

horr@fntimes.com

기사입력 : 2025-12-03 07:27

◇ Samsung Electronics pursues 'one-stop internalization' with self-sufficiency
◇ SK leverages 'collaboration' with TSMC, Hanmi Semiconductor, and others

Created with Gemini

Created with Gemini

이미지 확대보기
[Korea Financial Times, Gwak Horyung] In the High Bandwidth Memory (HBM) market, which has emerged as the core of memory performance in the AI era, Samsung Electronics and SK Hynix have entered a speed competition with diametrically opposite strategies.

From base die to back-end processes, Samsung Electronics plans to secure technological superiority through an internalization strategy encompassing foundry and equipment. In contrast, SK Hynix is leveraging an open collaboration model that mobilizes the entire external ecosystem including TSMC and Hanmi Semiconductor, leading with verified quality and speed.

HBM plays the role of quickly delivering necessary data to GPUs responsible for AI computations within a certain time. No matter how fast the GPU calculates, bottlenecks occur if data is not delivered on time. HBM is a structure that innovatively increases data processing speed by vertically stacking multiple DRAM chips.

Individual DRAM chips that make up HBM are called 'core dies.' The chip located at the bottom layer is specifically designated as the 'base die (logic die).' The base die performs the critical role of directly connecting to the GPU, interpreting signals, and delivering commands to upper core dies.

Until HBM3E (5th generation), memory companies produced base dies using their own DRAM processes, but from HBM4 (6th generation), foundry (contract manufacturing) methods utilizing cutting-edge fine processes are being introduced to improve speed and power efficiency.

SK Hynix is entrusting HBM4 base die production to Taiwan's TSMC. This is because the company judged that process optimization collaboration with TSMC is essential to meet performance requirements demanded by customers. SK Hynix is known to plan to utilize TSMC's 3-nanometer (nm) and 12nm processes. Standard products for servers will be manufactured with 12nm, while customized products such as Nvidia premium GPUs and Google's custom chips (TPU) are expected to be produced with 3nm processes.

In contrast, Samsung Electronics produces HBM4 base dies using its own foundry 4nm process. Although negotiations with customers over price and volume are variables, if Samsung Electronics secures prices similar to SK Hynix, it can create a relatively higher profit structure through internalization.

Samsung Electronics, a latecomer in HBM, has a strategy of securing customers by leveraging technological capabilities. Samsung Electronics applies the 1c (6th generation) process to HBM4 DRAM.

This is choosing a process one generation ahead of SK Hynix, which maintains the 1b (5th generation) process applied to existing HBM3E for HBM4 as well.

The HBM4 performance disclosed by the two companies in October also shows differences.

Samsung Electronics specifically presented data transfer (I/O) speed and bandwidth at maximum 11Gbps and 2.8TB per second, respectively. SK Hynix only disclosed minimum levels as 'above 10Gbps' and 'above 2.6TB.' This means that even the minimum standards meet Nvidia's enhanced performance requirements.

An industry official said, "For Samsung Electronics, securing yield is key," adding, "As a leading HBM company, SK Hynix's intention to prioritize stable supply is evident."

Packaging (back-end process) strategies also differ significantly between the two companies. The core back-end process in HBM is a type of bonding process that connects stacked DRAMs. From HBM4 20-stack and above, hybrid bonding is required for yield stabilization, but previous generations are likely to maintain existing technologies.

SK Hynix uses the 'MR-MUF process' that fills the space between chips with liquid protective material after soldering. The TC (thermal compression) bonder equipment that temporarily fixes chips to substrates and chips to chips just before this process is supplied by Hanmi Semiconductor, Hanwha Semitech, and others. Since even slight misalignment causes defects, TC bonders for HBM4 are known as expensive equipment costing approximately KRW 4 billion per unit.

If SK Hynix can be likened to 'applying wood glue,' Samsung Electronics can be compared to 'attaching double-sided tape.' This is the 'TC-NCF' method that inserts film-type protective material between layers and proceeds with thermal compression. Samsung Electronics mainly receives TC bonder supplies from its subsidiary SEMES.

Gwak Horyung (horr@fntimes.com)

데일리 금융경제뉴스 FNTIMES - 저작권법에 의거 상업적 목적의 무단 전재, 복사, 배포 금지
Copyright ⓒ 한국금융신문 & FNTIMES.com

가장 핫한 경제 소식! 한국금융신문의 ‘추천뉴스’를 받아보세요~

KFT Topic 다른 기사

1 Hyundai Steps Up U.S. Robot Lobbying as China Race Heats Up Hyundai Motor Group's lobbying expenditures targeting U.S. government agencies have surged sharply.What stands out is that the robotics industry has newly emerged on the lobbying agenda, which had previously focused on automotive industry issues such as auto tariffs, trade, EV tax policy and autonomous driving. Even Boston Dynamics, which oversees the group's robotics business, has been actively engaging with the U.S. Congress and government on robot-related legislation and national strategy.As the global humanoid robot market rapidly shifts from the R&D stage to a race for commercializati 2 LG's Koo Kwang-mo Wagers on Nvidia Alliance to Win the Physical AI Race LG Group Chairman Koo Kwang-mo's vision for "physical artificial intelligence (AI)" is taking concrete shape. Koo has chosen to combine LG's vertically integrated value chain — spanning components, finished products and platforms — with Nvidia's advanced AI "brain." By grafting Nvidia's proven AI platform onto the hardware value chain already built across LG's affiliates, the move appears designed to dramatically cut the time and cost of technology development. It is consistent with Koo's AI transformation strategy, which prioritizes fast execution over a flawless plan.Key Nvidia officials, 3 Naver's KRW 14 Trillion, 1GW AI Factory Deal Redraws Its Valuation Formula Naver is expanding beyond its roots as a search-and-advertising platform company into an artificial intelligence (AI) infrastructure provider, prompting a shift in how the market values the company. Naver plans to build a 1-gigawatt (GW)-class "AI Factory" backed by roughly KRW 14 trillion in investment and infrastructure procurement from Nvidia and Brookfield.According to industry sources on August 20, the business value of Naver's AI Factory operations is estimated at KRW 14.4 trillion — up KRW 3.3 trillion in about two months from the KRW 11.1 trillion valuation calculated in June, when Na
ad
ad
ad

한국금융 포럼 사이버관

더보기

FT카드뉴스

더보기
[그래픽 뉴스] ISA 대개편! 나에게 유리한 계좌는?
[그래픽 뉴스] 미국 증시 새로운 키워드 'MANGOS'
환전·로또·육아휴직까지 하반기부터 달라지는 제도 TOP11
[그래픽 뉴스] 은퇴후 30년 부모님 세대의 생존전략
[그래픽 뉴스] 퇴근 후 주차했는데 수익 발생? V2G의 정체

FT도서

더보기