SK hynix: AI inference needs full memory hierarchy optimization

SK hynix president Ahn Hyun said at SEDEX 2026 that the AI inference boom is creating a new memory bottleneck, with performance gaps between memory tiers and data movement now the key limit. The company is pursuing 3D Stacked DRAM, HBF with about half of HBM bandwidth at roughly 10x capacity, and custom HBM with logic on the base die.

• SK hynix sees a new bottleneck in memory tier gaps and data movement
• HBF targets about half of HBM bandwidth at roughly 10x capacity
• 3D Stacked DRAM stacks DRAM directly on GPUs and NPUs
• HBM will evolve into custom versions with logic on the base die