SK Hynix is accelerating its push into the AI infrastructure market by unveiling standard specifications for High Bandwidth Flash (HBF), a next-generation AI memory technology, in collaboration with SanDisk.
According to SK Hynix on the 4th, the company revealed the HBF standard specifications jointly developed with SanDisk at 'FMS (Future of Memory and Storage) 2026,' which opened on the same day (local time) in Santa Clara, California. This marks the first tangible result released approximately six months after the two companies launched the HBF Consortium in February of this year, following the start of standardization cooperation last August. HBF is a new memory tier positioned between HBM and SSDs. While it allows for the rapid exchange of data like HBM, it enables significantly increased capacity by utilizing NAND.
With the rapid increase in data that needs to be handled due to the spread of AI inference, it is garnering attention as a technology capable of simultaneously capturing both bandwidth and capacity scalability. The standard specifications unveiled this time support a maximum capacity of 512GB based on a structure that stacks NAND dies in 8 and 16 layers. 0 TB/s. By adopting UCIe (Universal Chiplet Interconnect Express), a semiconductor chip connection standard, compatibility with various processors, including GPUs and CPUs, has also been secured.
The standard specifications were released through the Open Compute Project (OCP), the world's largest open data center technology collaboration, and operate as an open standard that can be utilized by the entire industry, rather than specific companies. Currently, Google and the AI semiconductor company Tentorrent are participating in the HBF consortium, and SK Hynix plans to expand the ecosystem based on this foundation. SK Hynix will present changes in memory structure for the AI era at FMS (Future of Memory and Storage) 2026.
On the 4th, the first day of the event, Kim Cheon-seong, Head of Solution Development, and Kang Wook-seong, Head of Next-Generation Product Planning, will deliver a joint keynote speech on the theme of "Implementing Hierarchical Memory-Based AI Infrastructure in the Era of Agentic AI. " As a single type of memory alone is insufficient to meet the performance demands of AI systems, the necessity of a 'Tiered Memory' architecture, which utilizes various types of memory hierarchically, will be introduced. On the 6th, a panel discussion featuring representatives from Google DeepMind and SanDisk will also be held.
Under the theme "Crossing the Memory Wall with HBF," the discussion will cover changes in the memory structure of AI infrastructure and the role of HBF. At this exhibition, SK Hynix is introducing for the first time the wafers and products of its 10th Generation (V10) 375-layer 4D NAND, which is currently undergoing smooth development. 5-fold improvement in performance-to-power ratio compared to the previous generation.
SK Hynix plans to begin mass production of high-performance, high-capacity eSSDs utilizing this technology early next year. Division Head Kim Cheon-seong stated, "Through HBF, we will contribute to the realization of a new architecture that expands the boundaries between memory and storage and enhances overall system efficiency. " Reporter Kim Se-hyung fax123@sportschosun.com
This article has been translated by GripLabs Mingo AI.
