[Edaily Reporter JAEMIN SONG ] SK hynix has unveiled a “tiered memory” solution that combines not only High-Bandwidth Memory (HBM) but also DRAM, High-Bandwidth Flash (HBF)—the next-generation NAND—and enterprise solid-state drives (eSSDs). As the volume of data that artificial intelligence (AI) must process surges, this strategy aims to enhance performance and cost efficiency by connecting various types of memory according to their specific purposes.

SK hynix announced on the 7th that it participated in “FMS 2026,” a memory and storage exhibition held in Santa Clara, California, from the 4th to the 6th (local time), where it showcased its next-generation AI memory technologies and products.

During the keynote speech on the first day of the event, Kim Cheon-sung, Head of the Solution Development Division at SK hynix, and Kang Wook-sung, Head of Next-Generation Product Planning, served as co-speakers. Under the theme “The Era of Agent AI: Efficient Implementation of AI Infrastructure Based on Hierarchical Memory,” they explained the future direction of memory architecture evolution.

Hierarchical memory is a structure that arranges and organically connects memory types with different speeds, capacities, and costs according to their intended uses. Rather than storing all data required for AI computations in relatively high-cost HBM, this approach distributes data appropriately across HBM, DRAM, HBF, and SSDs, and quickly moves it as needed. This reduces data movement while simultaneously improving the performance, cost-effectiveness, and power efficiency of AI infrastructure.

Division Head Kim stated, “As AI expands from training-centric applications to large-scale inference and, ultimately, to agentic AI, infrastructure competitiveness is no longer determined solely by the performance of individual memory products.” He added, “Overall efficiency will depend on how each memory tier—including HBM, DRAM, NAND-based HBF, and SSDs—is positioned and interconnected to minimize data movement.”

He went on to emphasize, “Going forward, the ability to implement an optimized architecture that encompasses all memory tiers will be the key to competitiveness in AI infrastructure.”

The product SK hynix has identified as a new memory tier between HBM and SSDs is HBF. Like HBM, HBF is a next-generation NAND solution that utilizes through-silicon via (TSV) technology and logic processes to simultaneously deliver the high capacity of existing NAND-based storage devices and data transfer performance comparable to DRAM.

Lim Ui-cheol, Head of Solution AT at SK hynix, discussed the development direction and technical potential of HBF during a panel discussion held on the final day of the event alongside Google, SanDisk, and others. Lim stated, “HBF has the potential to establish itself as a new memory tier between HBM, an ultra-high-speed memory, and SSDs, which are high-capacity storage devices.” SK hynix is promoting joint hardware and software design and technical standardization with its key partners.

At the exhibition booth, the company showcased its product lineup for hierarchical memory all in one place. A model of NVIDIA’s next-generation AI accelerator, “Vera Rubin,” was displayed side-by-side with a model of HBM4, alongside the HBM4E 48 gigabyte (GB) 12-layer module, the HBM4 48GB 16-layer module, the HBM3E 36GB 12-layer module, and the SOCAMM2 memory module for AI servers.

In the NAND sector, the company unveiled a 375-layer 4D NAND utilizing wafer-bonding technology. This product is characterized by a 2.5-fold increase in performance per watt compared to the previous generation. SK hynix plans to begin mass production of high-performance eSSDs based on the 375-layer NAND in the first half of next year.

The company will also begin supplying samples of high-capacity eSSDs based on Quad-Level Cell (QLC) technology—which offer approximately a 55% performance improvement over existing 176-layer products—to major cloud service providers (CSPs) starting this month. This move is aimed at addressing the input/output performance, high bandwidth, and large-capacity storage demands required by AI data centers.

The company also demonstrated next-generation memory technology based on Compute Express Link (CXL). The company unveiled technology that efficiently stores and reuses AI agents’ key-value (KV) caches using “pooled memory”—where multiple servers and graphics processing units (GPUs) share a single memory pool—and “CMM-Hybrid,” which combines DRAM and SSDs. This approach distributes the KV cache—which grows as conversation and task histories expand—across CXL-based memory to boost AI inference speed and memory utilization.

Manager Kang stated, “SK hynix possesses a portfolio that spans all layers of AI infrastructure, from HBM to CXL-based memory and eSSD,” adding, “SK hynix’s role will be to work with customers to identify and provide the optimal solutions they need, starting from the product design stage.”