The Bottleneck in AI Hardware
The next-generation GPUs for artificial intelligence (AI) being designed by leading companies like NVIDIA, AMD, and Huawei are indeed remarkable technological achievements. However, their performance is heavily dependent on the memory chips that accompany them. The situation is so critical that these advanced GPUs often find themselves in a waiting game, forced to pause until the memory supplies the necessary data to continue processing.
Enter HBM4e Memory Chips
High Bandwidth Memory (HBM4e) chips are poised to eliminate this bottleneck in AI hardware once and for all. The leading designers of this type of integrated circuits—Samsung, SK Hynix, and Micron Technology—are actively developing their own HBM4e solutions. It is anticipated that Samsung and SK Hynix will deliver initial samples to their clients during the second half of 2026, whereas Micron plans to release theirs by 2027.
Market Dynamics and SK Hynix’s Strategic Decisions
Currently, SK Hynix dominates the HBM market with a share close to 70%, while Samsung and Micron share the remaining 30%. Maintaining this lucrative market position will require SK Hynix to produce its HBM4e chips using cutting-edge lithography technology. In this context, they are exploring a partnership with TSMC to manufacture the core of these memories on its advanced 3 nm node, as stated by DigiTimes Asia.
Why TSMC?
SK Hynix has multiple reasons for handing over the production of its HBM4e chips to TSMC. Firstly, TSMC is already involved in manufacturing the GPUs designed by SK Hynix’s primary clients. Thus, having TSMC also produce the memory simplifies the final assembly process using advanced COWOS (Chip-on-Wafer-on-Substrate) packaging.
Moreover, HBM4e chips require extremely small and fast transistors for optimal performance. SK Hynix recognizes that TSMC has more refined integration technologies than they currently possess. Furthermore, these new memories will not just store data; they will also perform basic operations on the data before passing it to the GPU, behaving somewhat like processors themselves.
The TSMC Challenge
However, this strategic alliance comes with challenges. TSMC’s N3 node is currently overwhelmed with demand from clients like NVIDIA and Apple. TSMC has been facing significant difficulties in meeting the requirements since it began manufacturing 3nm chips, struggling with its second-generation technology, N3E, to enhance yield per wafer.
The Future Outlook
The N3P lithography, the next generation after N3E, offers improved transistor density and performance while reducing power consumption. Yet, even with these advancements, the supply of wafers remains a pressing issue. Thus, the uncertainty around how TSMC will accommodate SK Hynix’s entry into this saturated node remains a significant concern.
In conclusion, while the innovations in HBM4e memory represent a thrilling advancement in AI hardware, the interplay between supply, demand, and technological capabilities will be crucial in shaping the future landscape of this rapidly evolving industry.