Samsung Electronics and SK Hynix are charting different paths in the next market for high-bandwidth memory (HBM). Samsung's strategy focuses on a 'custom-tailored' approach, directly integrating memory onto customer artificial intelligence (AI) chips, while SK Hynix is pursuing a 'shared standard' that can be utilized by multiple clients.
On August 5, the semiconductor industry revealed that both companies showcased their next-generation AI memory technologies at the 'FMS 2026' event in Santa Clara, California. The technological directions for capturing market leadership post-HBM have become distinctly divided.
Samsung's leading technology is zHBM. Unlike traditional HBM, which is placed next to AI accelerators, zHBM stacks memory vertically on top of the AI accelerator. This reduces the distance between the processor and memory, increasing data transfer speeds and reducing power consumption.
Samsung claims that zHBM can deliver up to eight times the performance compared to HBM5. It also boasts a power-to-performance ratio that is three times higher and can reduce thermal resistance by more than half. As AI models grow larger, the data transfer bottleneck increases, and Samsung's solution is to place memory closer to the processor.
Custom design is also a key aspect. Samsung explained that it can incorporate customer-specific intellectual property (IP) into the interlayer between the memory and AI accelerator. This means the structure can be adjusted to meet the desired memory capacity, accelerator performance, and power conditions of the client.
In NAND technology, Samsung emphasized its 3D strategy, unveiling the 10th generation V-NAND, V10 BV-NAND, which has over 400 layers. By applying wafer bonding technology, it has increased density by approximately 58% compared to the previous generation. The zNAND-O for on-device AI was also introduced.
In contrast, SK Hynix is focusing on standardization. It unveiled the first standard specification for high-bandwidth flash (HBF) in collaboration with SanDisk. HBF is a next-generation storage technology that, like HBM, allows for rapid data transfer while increasing capacity based on NAND.
HBF serves as a new layer between HBM and solid-state drives (SSDs). As AI inference becomes more widespread, data will need to be stored more frequently and accessed more often. Relying solely on HBM would create capacity burdens, while SSDs alone would lack sufficient speed. SK Hynix aims to fill this gap with HBF.
The new specification is based on a maximum capacity of 512GB, with bandwidth divided into three tiers ranging from approximately 0.4TB to 3.0TB per second. It adopts the open chiplet interface UCIe for connections with processors, allowing for compatibility with various chips, including GPUs and CPUs.
SK Hynix also introduced its 10th generation 375-layer 4D NAND at FMS, which offers 2.5 times the power-to-performance ratio compared to the previous generation. The company plans to begin mass production of high-performance, high-capacity eSSDs based on this technology early next year.
A semiconductor industry insider noted, 'The post-HBM competition is a battle over who can dominate the AI system architecture rather than just product specifications. Samsung targets optimization for individual customers, while SK Hynix aims for an open ecosystem to address AI memory bottlenecks.'
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.