DEEP RESEARCH · SEMICONDUCTORS/HBM
HBM-Centric Scenario and Korea’s Semiconductor Strategy
A structured note from Professor Jung-Ho Kim’s video on HBM technology, AI bottlenecks, and the future of Korea’s semiconductor industry
0. Bottom line first
This is a video note I saved to watch later. The key point is that Samsung is developing its own HBM-centric chip, while SK hynix is developing within an ecosystem together with NVIDIA and TSMC. The note carries the view that in the AI era, HBM, which supports GPU compute speed, may become even more important than the GPU itself.
1. Where HBM Technology Started
Official fact: According to the source note, Professor Jung-Ho Kim of KAIST's School of Electrical Engineering has researched the foundations of HBM technology since 2000 and worked with SK hynix on actual design. His lab is known as the only university lab in the world researching HBM and has produced many HBM specialists. The video timestamps noted are [01:20], [01:48], and [01:57].
HBM technology was developed to overcome the limits of improving memory density and speed. The explanation continues that, under the expectation that Moore's Law would reach its limits, the focus shifted to stacking methods that increase memory capacity and bandwidth. The related timestamps are [06:15] and [06:26].
2. Why HBM Matters More in the AI Era
Official fact: The source note says the rise of AI technologies such as generative AI is increasing the importance of HBM. HBM plays a key role in solving the memory bottleneck where memory cannot keep up with GPU compute speed. The related timestamps are [13:33], [15:30], and [17:05].
Interpretation: Professor Kim emphasizes that in the AI era, HBM is much more important than GPUs and that Korea’s semiconductor industry should focus on HBM. The timestamps in the source note are [19:35] and [19:56].
Memory bandwidth
The key constraint in the AI era is presented as memory failing to keep pace with GPU compute speed.
HBM
HBM is described as the technology axis that eases the GPU bottleneck through high bandwidth.
Korea’s focus area
The view is that Korea’s semiconductor industry should build on memory capability and focus on HBM.
3. Different Scenarios for Samsung and SK hynix
The first memo in the source distinguishes Samsung and SK hynix. Samsung is developing its own HBM-centric chip, while SK hynix is developing within an ecosystem together with NVIDIA and TSMC.
| Company | Direction noted in the source | Question for investors |
|---|---|---|
| Samsung | Developing its own HBM-centric chip | Which customers and ecosystem will the proprietary chip strategy connect to? |
| SK hynix | Developing in the ecosystem with NVIDIA and TSMC | Can its collaboration advantage inside the AI semiconductor ecosystem continue? |
4. Future Roadmap and Korea’s Semiconductor Industry
Official fact: The KAIST lab is noted as planning to announce a roadmap for HBM 4, 5, 6, and 7 architectures. Future HBM technology is expected to evolve not only through more 3D stacking layers but also by expanding HBM chips sideways. The timestamps are [24:22], [27:57], and [29:00].
The note says Korea must secure leadership in the HBM market based on its memory semiconductor capability, and that this is directly tied to national survival. In the AI era, HBM-centric computing, which 3D-packages CPU, GPU, and other components around HBM, will become important, and training convergence-type talent is urgent. The related timestamps are [21:31], [22:06], [44:12], [44:50], and [48:56].
5. Core Takeaway From My Note
- HBM should be viewed not as a peripheral technology supporting GPUs, but as a core axis for solving AI-era bottlenecks.
- Samsung and SK hynix are taking different approaches to HBM.
- Korea’s semiconductor industry needs to convert memory capability into HBM leadership.
- Professor Kim emphasizes that industry-academia collaboration and active government investment are necessary.
Sources
- Original Naver Blog post: https://m.blog.naver.com/PostView.naver?blogId=star_of_self&logNo=223874365291