返回资讯中心
外部精选
软件工程
#开发者实时流

SK Hynix Unveils Next-Generation Memory Roadmap to Address AI Bottlenecks

SK hynix is accelerating the development of next-generation memory technologies to address growing bottlenecks in artificial intelligence (AI) infrastructure as demand for AI inference continues to surge. The company pl…

Hacker News 最新peter_d_sherman2026/10/15 · 2026/10/1514 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

SK hynix is accelerating the development of next-generation memory technologies to address growing bottlenecks in artificial intelligence (AI) infrastructure as demand for AI inference continues to surge. The company plans to strengthen its high-bandwidth memory (HBM) leadership while developing 3D Stacked DRAM and High Bandwidth Flash (HBF) to improve system performance and energy efficiency.

Ahn Hyun, Chief Development Officer (CDO) of SK hynix, presented the company's technology roadmap during a keynote address at SEDEX 2026, South Korea's largest semiconductor exhibition, on Oct. 14. Under the theme “The Evolution of Memory Technology in the AI Era,” Ahn highlighted the need to optimize the entire memory hierarchy rather than focus solely on improving individual products.

Ahn said AI inference demand is expanding rapidly across three dimensions: task complexity, context length and inference request volume. These factors reinforce one another, driving a sharp increase in demand for AI computing infrastructure. SK hynix estimates that the scale of AI infrastructure performance requirements could grow approximately 20-fold by 2030 compared with 2026.

As AI models process increasingly large volumes of data, bottlenecks are emerging not only in computing performance but also in data storage and transfer. Key-value (KV) caches, which store previously processed information for reuse, require significantly more memory capacity as context lengths increase.

Managing these large datasets requires information to be distributed across multiple memory tiers, including HBM, DRAM and solid-state drives (SSDs). However, moving data between these layers can increase latency and power consumption.

Ahn noted that the traditional “memory wall,” which refers to the performance gap between processors and memory, is evolving in the AI era. Differences in performance across memory tiers and the energy costs associated with data movement are becoming increasingly important constraints.

SK hynix therefore plans to develop an integrated memory architecture that connects different tiers more efficiently, supporting the continued expansion of AI infrastructure.

The company will continue advancing HBM technology, improving capacity, bandwidth and energy efficiency across generations from HBM1 to HBM4E. It is also expanding its focus to custom HBM, which incorporates memory controllers and customer-specific logic into the base die to better match the requirements of individual AI systems.

Beyond HBM, SK hynix is developing 3D Stacked DRAM, which vertically stacks DRAM on top of logic chips such as graphics processing units (GPUs) and neural processing units (NPUs). By reducing the physical distance between computing components and memory, the technology aims to deliver data more quickly while improving performance and energy efficiency.

The company also introduced HBF as a potential solution for the growing capacity demands of AI inference. HBF applies through-silicon via (TSV) stacking and high-speed input/output technologies developed for HBM to NAND flash memory. Positioned as a new tier between HBM and DRAM, HBF is designed to combine high bandwidth with substantially greater storage capacity.

SK hynix is targeting bandwidth at approximately half that of HBM of the same generation while providing around 10 times the capacity. The company expects HBF to help reduce reliance on DRAM and SSD access, improving overall system performance and energy efficiency.

Ahn also highlighted the suitability of NAND-based HBF for AI inference workloads, which frequently involve repeated data reads while generating relatively fewer write operations. SK hynix plans to pursue early standardization of HBF and strengthen collaboration with global partners to establish a broader ecosystem for the technology.

In addition, the company is exploring technologies such as Compute Express Link (CXL) and Storage-Next to diversify memory and storage architectures and better accommodate different AI workloads.

At SEDEX 2026, SK hynix is also showcasing a broad portfolio of memory and storage solutions spanning AI infrastructure, personal computers and mobile devices. Its exhibition booth, the largest at the event, adopts the theme “An Unstoppable AI Journey” and is designed around a running track that guides visitors through the company's technology and product portfolio.

Featured server memory products include the 256GB 3DS RDIMM, designed to deliver high capacity for AI workloads; the 128GB MRDIMM, which offers high bandwidth; and the 64GB RDIMM, which uses the company's sixth-generation 10-nanometer-class (1c) process.

The company is also presenting enterprise SSDs, including the PEB210 E1.S 9.5mm and PEB210 M.2 2280, which support direct liquid cooling, as well as the PS1101 E3.S 7.5mm, its first SSD developed using quad-level cell (QLC) technology.

For PC and mobile applications, the lineup includes DDR5 CSODIMM, GDDR7, LPDDR5X and LPDDR6 memory products, alongside PC SSDs and mobile NAND solutions such as UFS 4.1 and UFS 5.0.

The exhibition also features models illustrating HBM and HBF architectures, displays tracing the development of HBM, and an NVIDIA Vera Rubin superchip. An HBM wafer bearing NVIDIA CEO Jensen Huang's signature highlights the companies' partnership.

Through its latest technology roadmap and exhibition, SK hynix aims to extend its memory expertise beyond HBM and develop a more integrated memory ecosystem. By combining higher-performance products with optimized memory hierarchies, the company seeks to address the growing data processing demands of the AI era.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →