SK hynix CEO Says HBM from 2025 Production Almost Sold Out

[btarunr]

SK hynix held a press conference unveiling its vision and strategy for the AI era today at its headquarters in Icheon, Gyeonggi Province, to share the details of its investment plans for the M15X fab in Cheongju and the Yongin Semiconductor Cluster in Korea and the advanced packaging facilities in Indiana, U.S.

The event, hosted by theChief Executive Officer Kwak Noh-Jung, three years before the May 2027 completion of the first fab in the Yongin Cluster, was attended by key executives including the Head of AI Infra Justin (Ju-Seon) Kim, Head of DRAM Development Kim Jonghwan, Head of the N-S Committee Ahn Hyun, Head of Manufacturing Technology Kim Yeongsik, Head of Package & Test Choi Woojin, Head of Corporate Strategy & Planning Ryu Byung Hoon, and the Chief Financial Officer Kim Woo Hyun.



The opening speech by Kwak was followed by Justin Kim's presentation on the company's vision for AI memory, Choi Woojin's introduction to SK hynix's core technology for HBM and the plan for the advanced packaging facilities in the U.S., and Kim Yeongsik's sharing of the plans for the M15X fab and the facilities for the Yongin Cluster. A question-and-answer session also followed.

Below are the key messages from the presentations at the event.

• Company forecasts a fast expansion of AI technology into a wider range of on-device applications such as smartphones, PCs, and automobile from data center now
• Demand for ultra-fast, high-capacity and low-power memory products for AI applications expected to show an explosive increase
• Company has industry's best technologies for various products including HBM, TSV-based high-capacity DRAM, and high-performance eSSD
• SK hynix ready to provide industry's best customized memory solutions to customers through strategic collaboration with global business partners
• On production side, HBM from 2024 output already sold out, while that from 2025 volume almost sold out
• On HBM technology side, company planning to provide samples of 12-high HBM3E with industry's best performance in May, enablestart of mass production in 3Q
• Company aims for a qualitative growth through better cost competitiveness, a higher profitability with an increase in sales of value-added products
• Plan is to continue to improve financial soundness by raising level of cash holding through a flexible investment response in accordance with changing circumstances for demand
• Company committed to contributing to domestic economy, helping advance Korea's position as an AI memory powerhouse by growing into a trusted customer, stable company not swayed by business circumstances in AI era

Total volume of data generated globally in AI era forecast to jump to 660 zettabytes in 2030 from 15 ZB in 2014.

• Changes in semiconductor industry in AI era require a paradigm shift and a new approach to industry
• Memory is the clear answer to the question of who can provide a differentiated value in AI era
• Memory at the heart of the data-centric AI era as virtuous cycle of storage, accumulation, and reproduction of data required
• A sharp rise in sales ratio of AI memory expected with portion of AI memory led by HBM and high-capacity DRAM module forecast to grow to 61% in 2028 in terms of value from approximately 5% in 2023
• Company has maintained technological leadership in various AI applications
• In DRAM space, company mass producing HBM3E and modules with ultra-high capacity of more than 256 GB, while having commercialized world's fastest LPDDR5T
• Company a top provider of AI memory also in NAND space as sole supplier of QLC-based SSD of more than 60 TB
• Development of next-generation products with improved performance underway
• Company planning to introduce innovative memory such as HBM4, HBM4E, LPDDR6, 300 TB SSD, CXL Pooled Memory Solution, and Processing-In-Memory

Core Technology for HBM and Advanced Packaging Facilities in the U.S. by Choi Woojin

• SK hynix's proprietary MR-MUF is a core technology for HBM packaging
• Views that MR-MUF will face technological challenges in higher stacking incorrect as seen in SK hynix's successful mass production of 12-high HBM3 with Advanced MR-MUF technology
• MR-MUF lowers pressure from chip stacking to 6% level, raises productivity by 4 times by shortening time required for process, while improving heat dissipation by 45% VS previous technology
• Advanced MR-MUF recently introduced by SK hynix improves heat dissipation by 10% by adopting a new protective material, while maintaining existing advantages of MR-MUF
• Advanced MR-MUF, which adopts high-temperature, low-pressure methodology known for excellent warpage control, an optimal solution for high stacking and development of technology to realize 16-high stacking underway
• Company plans to adopt Advanced MR-MUF for realization of 16-high HBM4, while preemptively reviewing Hybrid Bonding technology
• Separately, company announced last month a plan to build advanced packaging facilities for AI memory in West Lafayette, Indiana
• Mass production of AI products such as next-generation HBM from Indiana fab to start in 2H 2028
• Indiana is center of the Silicon Heartland, a semiconductor ecosystem centered in the Midwest
• SK hynix to strengthen collaboration with customers, competitiveness in AI space from Indiana, while fostering talent through cooperation with R&D centers to contribute to future of AI industry

Investment Plans for M15X in Cheongju and Yongin Semiconductor Cluster by Kim Yeongsik

• Company plans to build M15X fab in Cheongju, North Chungcheong Province, as a rapid rise in demand for AI memory requires capacity expansion before operation at Yongin Cluster starts
• M15X, a two-story fab that accommodates various processes for HBM production including EUV, expected to maximize production efficiency thanks to adjacency to M15, which has been expanding TSV capacity
• Construction of M15X has started last month for completion in November 2025, mass production from 3Q26
• Separately, company also plans to build the Yongin Semiconductor Cluster, home to fabs, complex for business partners and infrastructure
• Four fabs to be ready in order, while suppliers of materials, components and equipment to move in to build semiconductor ecosystem through collaboration with SK hynix
• Process to form the site underway with progress rate of leveling the ground for 1st fab standing atapproximately 42%
• Construction of 1st fab to start in March 2025 for completion in May 2027
• Mini fab to be located within the cluster to help domestic business partners with valuable ideas to develop new technology and win competitiveness by offering opportunities to test under practical circumstances for mass production
• Cleanroom and human resources to be provided by SK hynix for free, while central and local governments will support investment and operation of equipment
• SK hynix to work towards strengthening domestic ecosystem and Korea's leadership in semiconductor industry through Yongin Cluster

View at TechPowerUp Main Site

 

[Panther_Seraphin]

My money is on this being why the high end RX8xxx GPUs were canned before hitting production.

 

[bonehead123]

Sooooo to summarize:

Everything they make from this point forward will be much more expensive, especially consumer GPU's, since somebody has to pay for all that AI growth & expansion....and it AINT gonna be them

hint hint... it will be us little people, those who can least afford it !

 

[Panther_Seraphin]

HBM is a fantastic product and is only really finding it niche in recent times due to the AI boom and the increases it has seen from HBM1 up to HBM3e and I look forward to it coming down to consumers soon but it will definately be 2-3 generations before we really see it back into consumer cards/cpus/memory soloutions.

Current HBM3e offerings would offer DOUBLE the bandwidth of the already insane bandwidth Vega VII cards had and latest examples are showing 36GB of Storage per die. So something like a 2Tb/s Bandwidth with 72Gb of VRAM for two dies.........YES PLEASE!!!

 

[Denver]

My money is on this being why the high end RX8xxx GPUs were canned before hitting production.

What does the shortage of HBM have to do with gaming GPUs, which use GDDR6?

I think AMD just chose to allocate production capacity to something more profitable (mi300). They also have a billion-dollar HBM3 contract from Samsung, so gaming GPUs will become scarce eventually. Almost all modern chips use 4/5nm

 

[Panther_Seraphin]

What does the shortage of HBM have to do with gaming GPUs, which use GDDR6?

I think AMD just chose to allocate production capacity to something more profitable (mi300). They also have a billion-dollar HBM3 contract from Samsung, so gaming GPUs will become scarce eventually. Almost all modern chips use 4/5nm

The mid range GPUs are meant to be using GDDR6, With the proclivity of HBM in the Datacenter recently and the fact AMD has already used HBM previously in Vega VII and Vega 56/64 products makes me think they may have taken the risk again putting it on to their RX8xxx gen of graphics cards at the tippy top parts but were caught out when the availability/costing for HBM didnt come in at where they were hoping.

Think of an RX7900xtx die with 2-4 HBM dies of memory vs GDDR

 

[Denver]

The mid range GPUs are meant to be using GDDR6, With the proclivity of HBM in the Datacenter recently and the fact AMD has already used HBM previously in Vega VII and Vega 56/64 products makes me think they may have taken the risk again putting it on to their RX8xxx gen of graphics cards at the tippy top parts but were caught out when the availability/costing for HBM didnt come in at where they were hoping.

Think of an RX7900xtx die with 2-4 HBM dies of memory vs GDDR

I highly doubt anyone at AMD has even considered the possibility, given that some HBM3 chips are pricier than entire gaming GPUs. AMD likely learned their lesson from their experiences with HBM2.

AMD has consistently aimed for maximum efficiency in GPU performance per area and production costs. They've even dabbled in MCM designs to enhance their offerings, optimizing costs Ryzen style. However, the rumored MCM with multi GCDs might have been scrapped due to technical challenges or AMD's reluctance to invest further for potentially limited gains. This could potentially free up additional capacity to ride the AI wave with the Mi300X and its variations.

 

[LabRat 891]

The mid range GPUs are meant to be using GDDR6, With the proclivity of HBM in the Datacenter recently and the fact AMD has already used HBM previously in Vega VII and Vega 56/64 products makes me think they may have taken the risk again putting it on to their RX8xxx gen of graphics cards at the tippy top parts but were caught out when the availability/costing for HBM didnt come in at where they were hoping.

Think of an RX7900xtx die with 2-4 HBM dies of memory vs GDDR

While I adore this sentiment, I don't think it's realistic.

HBM is going into high-profit products, like the MI APUs and AI/MI Accelerators.

W/ 'big Navi 4x' being supposedly 'real' and cancelled, I'd bet RX 8k is going to be a re-do of the HD 8k series: Mostly OEM sales.
If nVidia delays or 'restricts supply' of their next gen card, AMD can happily chug along w/ price-cut RX7ks for a bit.

AMD's next big GPU series will be RX 9000s. How do I 'know' this?
AMD's clearly been working towards 'market pairing' the models of their CPUs and GPUs. Ex. Ryzen 8k are mainstream APUs, a lot like the Ryzen 4000series.
IMO, Radeon RX 8000 will be 'mainstream' high-yield RDNA 3.5.
RX 9000 might come closer to Panther and I's dreams, though.

 

[Minus Infinity]

HBM will never be a thing in desktop gpus. HBM costs a bomb to make and even if supply weren't an issue it's simply too expensive and large for a desktop gpu.

 

[Panther_Seraphin]

It was done however previously in 2 generations of GPUs from AMD even with costs being less than 50% of top end cards now.

I suspect when it does return It will return only on Halo/Flagships only and will NOT filter down for quite a while. So 5090s/9900xt/xtx would be the only ones with this capability.