SK Hynix Unveils 'Tiered Architecture' to Divide Memory Roles in the AI Era

Single memory has reached its limits. On the 7th, SK Hynix unveiled a 'tiered memory' strategy where data is processed according to the strengths of different memory types, ranging from high-speed memory to storage space. As AI recently evolves beyond simple learning to the 'AI agent' stage—where it thinks independently—the company introduced technologies that stage memory products like HBM, DRAM, NAND, and SSDs to efficiently handle surging data.

SK Hynix explained that the core of tiered memory lies not just in improving individual product performance, but in placing them in optimal positions to minimize data movement. In particular, HBF (High Bandwidth Flash), a next-generation NAND solution, was introduced as a new memory layer as a technical means to realize this.

At the exhibition hall, technologies that enable AI agents to efficiently store and reuse data, such as conversation context, were demonstrated using CXL-based Pooled Memory technology and CMM-Ax, which adds computational functions to memory.

Additionally, SK Hynix exhibited HBM4 and HBM4E products alongside NVIDIA's next-generation AI accelerator 'Vera Rubin,' confirming plans to provide optimal solutions to customers through 'co-design,' which involves designing hardware and software simultaneously.