Solutions to the 'Memory Wall' emerge for the AI Agent era, presenting contrasting strategies focused on 3D vertical stacking and NAND utilization.
Samsung Electronics and SK Hynix have unveiled their next-generation memory strategies targeted at the era of "AI Agents." As artificial intelligence evolves into a phase where it independently makes decisions and executes tasks, overcoming the "memory wall"—which fundamentally determines system performance—has emerged as a critical challenge. In response, Samsung Electronics has championed an integrated turnkey solution leveraging vertical integration, whereas SK Hynix has positioned open standards and ecosystem expansion as its primary strategic gambits.

Both industry leaders introduced advanced 3D stacking techniques and novel approaches to NAND flash utilization that extend beyond traditional High Bandwidth Memory (HBM). However, their execution paths diverge significantly: Samsung emphasized a "one-stop strategy" encompassing memory, foundry services, and advanced packaging, while SK Hynix centered its roadmap on building a collaborative AI memory ecosystem rooted in standardization.

These strategic visions were presented during the opening of "FMS 2026," a premier global memory conference held on August 4 (local time) at the Santa Clara Convention Center in California. Throughout the event, overcoming the AI memory wall remained the paramount topic of discussion. As AI transitions from answering simple queries to autonomously formulating and executing plans in agent form, the role of memory in processing and transferring data has become fully as vital as computational processing power.

Samsung Electronics: Vertical 3D Integration and Turnkey Solutions

Samsung Electronics introduced its breakthrough next-generation technology, "zHBM," which stacks memory directly on top of the AI accelerator. Unlike conventional HBM, where memory units are placed adjacent to the processor horizontally, zHBM stacks memory vertically directly over the processing unit, drastically shortening the physical distance data must travel. By transcending planar architectural limits through three-dimensional stacking, Samsung aims to simultaneously maximize performance and power efficiency.

According to Samsung Electronics, zHBM targets up to 8x higher performance and up to 3x improved power efficiency compared to HBM5, which is anticipated for mass production around 2028. To resolve heat dissipation challenges, Samsung focused on optimizing data transmission speeds and thermal management rather than simply maximizing stack height.

Additionally, Samsung unveiled "zNAND-O," a solution that integrates vertically stacked NAND flash directly with a Neural Processing Unit (NPU). Leveraging its comprehensive one-stop model—encompassing design, foundry manufacturing, and advanced packaging under a single roof—Samsung outlined a cohesive strategy to target the rapidly expanding on-device AI market.

SK Hynix: Tiered Memory Hierarchy and Open Ecosystem

In contrast, SK Hynix placed its primary emphasis on fostering an open ecosystem. The company unveiled its 3D stacked DRAM, dubbed "G0.5," designed to bridge the performance gap between ultra-fast but low-capacity SRAM and high-capacity HBM. By further segmenting the memory hierarchy in AI computing workflows, SK Hynix aims to dramatically enhance total system efficiency.

SK Hynix also introduced High Bandwidth Flash (HBF) for AI servers, introducing a tiered memory architecture that blurs the traditional line between memory and storage. Rather than relying exclusively on high-cost DRAM for ultra-large AI workloads, this architecture incorporates NAND flash to optimize both operational cost and system throughput.

SK Hynix highlighted a distinct approach in ecosystem development as well. The company publicly disclosed the HBF standard specifications—jointly developed with SanDisk and other partners—through the Open Compute Project (OCP), the world's largest open-source data center technology consortium. This move aims to build an inclusive AI memory ecosystem accessible to diverse industry players.

During his keynote address at the event, Kim Cheon-seong, Head of Solution Development at SK Hynix, stated: "As AI adoption accelerates rapidly, we have reached a pivotal moment that demands a total redesign of overall data processing architectures. Through HBF, we intend to expand the boundary between memory and storage and contribute to implementing new architectures that enhance efficiency across the entire system."

[Copyright (c) Global Economic Times. All Rights Reserved.]