SK Hynix proposes HBM and HBF hybrid for LLM inference

What happened

SK Hynix has proposed a hybrid memory architecture called H³, combining HBM and HBF on a single interposer to address KV cache capacity bottlenecks in large-scale LLM inference. This is a research paper presented at IEEE and addresses a technical challenge where models like Llama 4 can require multi-terabyte caches, exceeding current HBM capacity. The proposal is a long-term R&D concept with no immediate impact on production, supply, or pricing of existing memory components.

Impact and signal

Affected components and suppliers

Source and time

Source: Blocks & Files · 2/16/2026 · View original source