TrendForce: SK hynix becomes the first to launch HBM3e 16Hi product, raising the upper limit of bit capacity.

TrendForce reported that SK hynix recently revealed at the SK AI Summit 2024 event that they are developing HBM3e 16hi products, with each HBM chip having a capacity of 48GB, and they are expected to be sampled in the first half of 2025. According to the latest research from TrendForce, potential applications of this new product include ASIC developed by CSP (cloud service providers) and general purpose GPUs, with the possibility of increasing the bit capacity limit ahead of the mass production of HBM4 generation, earlier than the HBM3e generation. TrendForce also mentioned that in the past, HBM suppliers would release products with different stack layers in each generation, such as 8hi and 12hi for HBM3e generation, and planning 12hi and 16hi for HBM4 generation. Given that many companies have already planned to invest in HBM4 12hi after the second half of 2025, SK hynix's decision to introduce 16hi products for HBM3e can be attributed to the following factors: Firstly, TSMC's CoWoS-L packaging size will expand between 2026 and 2027, allowing more HBM chips to be integrated into each SiP (System-in-Package). Before the more challenging HBM4 16hi production, offering customers the HBM3e 16hi as a high bit capacity product choice can be beneficial. With the assumption of 8 HBM chips per SiP, HBM3e 16hi can increase the bit capacity limit to 384GB, surpassing NVIDIA Rubin's 288GB capacity. The transition from HBM3e to the HBM4 generation will double the number of IOs, increasing the operational bandwidth. This upgrade leads to larger die sizes, while the capacity per die remains at 24Gb. During the transition from HBM3e 12hi to HBM4 12hi, HBM3e 16hi can serve as an option for low IOs, smaller die sizes, and high bit capacity. SK hynix's HBM3e 16hi will use the Advanced MR-MUF stack process, which is easier to achieve high stack layers and operational bandwidth compared to the TC-NCF process. With the expectations of designing 16hi products for both HBM4 and HBM4e generations, SK hynix's early mass production of HBM3e 16hi will help accumulate production experience for this stack layer and accelerate the mass production schedule for HBM4 16hi. SK hynix also does not rule out the possibility of introducing a hybrid bonding process version of the 16hi products to expand the application audience requiring higher operational bandwidth.