SK Hynix unveils MARS rack-scale memory architecture, HBF capacity up to 14 times that of HBM, samples targeted for 2028

At the 2026 OCP Global Summit, SK hynix Senior Vice President Kim Ho-sik introduced MARS, a memory-centric rack-scale AI system featuring a memory-decoupled pooling design that boosts large-model first-token generation speed by up to 5x, addressing the pain points of wasted memory resources and over-provisioned GPUs in traditional servers. He also disclosed progress on HBF high-bandwidth flash: its bandwidth is about half that of HBM, but its capacity can reach 14 times that of HBM, filling the storage hierarchy gap between HBM and SSD and catering to AI inference KV cache requirements. The HBF specification is expected to release version 1.0 next year, with product sampling anticipated in early 2028.