SK hynix and SanDisk Launch High Bandwidth Flash Standard to Break AI Memory Wall
The new HBF specification bridges the performance gap between HBM and SSDs to scale AI inference.
SK hynix and SanDisk have introduced the first standard specifications for High Bandwidth Flash (HBF), a new storage tier designed to eliminate critical memory bottlenecks in AI systems. Unveiled at FMS 2026 on August 4, the technology combines the massive capacity of NAND flash with speeds that approach High Bandwidth Memory (HBM).
To ensure a broad industry ecosystem, the companies released the specifications through the Open Compute Project (OCP). The HBF standard supports capacities up to 512GB, utilizing 8-high and 16-high NAND stack configurations. Performance is divided into three distinct bandwidth grades, delivering speeds ranging from approximately 0.4TB/s to 3.0TB/s. To facilitate seamless integration with CPUs and GPUs, the standard adopts Universal Chiplet Interconnect Express (UCIe). The initiative has already attracted major industry players, with Google DeepMind and Tenstorrent joining as participating members of the HBF consortium.
The Tiered Memory Shift
This development addresses a growing crisis in AI infrastructure known as the "memory wall." As inference models scale, the volume of data required for processing has surged beyond the capabilities of current architectures. Until now, developers faced a binary choice: use HBM for extreme speed at a high cost and limited capacity, or use SSDs for massive storage at speeds too slow for real-time computation. HBF introduces a "Tiered Memory" solution, acting as an intermediate layer that allows massive datasets to remain accessible at speeds significantly higher than traditional flash storage.
Industry Implications
By transforming the memory hierarchy into a more granular spectrum, HBF could fundamentally reshape the total cost of ownership (TCO) for AI data centers. This architectural shift is critical for the commercialization of Agentic AI and other large-scale systems that require vast amounts of data to be processed with near-real-time latency. According to Kim Chun-sung, SK hynix Executive Vice President and Head of Solution Development, the goal is to expand the boundaries between memory and storage to enhance overall system efficiency.
Future Outlook
Beyond the HBF standard, SK hynix is advancing its underlying hardware with the unveiling of a 10th-generation (V10) 375-layer 4D NAND. This new NAND technology offers a 2.5x increase in power efficiency, with mass production of enterprise SSDs (eSSDs) scheduled for early 2027. While the HBF specifications are now public via the OCP, the industry will be watching for the first hardware implementations and the subsequent integration of these chiplets into next-generation AI accelerators.