SK Hynix, together with SanDisk, has unveiled the first standard specification for HBF (High Bandwidth Flash), a next-generation artificial intelligence (AI) memory technology. As the performance gap between High Bandwidth Memory (HBM) and Solid-State Drives (SSDs) widens with the proliferation of AI inference, the move is a strategic bid to secure leadership in the AI infrastructure market by staking a claim to the new standard that bridges these two memory tiers.

SK Hynix announced the HBF standard specification on August 4 (local time) at the ‘FMS (Future of Memory and Storage) 2026’ exhibition, which opened at the Santa Clara Convention Center in California. The specification was released through the OCP (Open Compute Project), the world’s largest open data center technology collaboration, positioning it as an open standard available to the entire industry rather than being limited to specific companies.

HBF is a new memory tier situated between HBM and SSDs. Its core purpose is to combine HBM-level high data processing performance with NAND flash-based large-capacity storage capabilities to reduce memory bottlenecks that occur during AI inference and large-scale data processing. As AI models become more sophisticated, the importance of a memory tier capable of rapidly supplying large volumes of data—not just high-speed computation—has grown, making HBF a focal point as a next-generation AI memory technology.

The new standard supports capacities up to 512GB based on a structure that stacks NAND dies in 8-high and 16-high configurations. Data transfer speeds are defined in three grades (Grade 1 to 3), ranging from approximately 0.4TB/s to a maximum of 3.0TB/s per second. The interface connecting HBF to processors adopts the industry-standard UCIe (Universal Chiplet Interconnect Express), designed for flexible application across various processors including GPUs and CPUs.

CategoryDetailsCapacity512GB (8-high and 16-high NAND die stacking)Bandwidth0.4TB/s to 3.0TB/s (three grades)Processor ConnectionUCIe (open chiplet interface)

Note: The standard specification includes system interfaces, electrical characteristics, reliability and packaging guidelines for HBF die stacks, and software usage guidelines.

HBF’s position in the memory landscape becomes clearer when viewed alongside existing memory and storage tiers.

Memory TierBandwidthCapacity (per stack/drive)InterfaceHBM3E (GPU-direct DRAM)Up to ~1.2TB/s per stackUp to 36-48GB per stack2,048-bit TSV (Through-Silicon Via)HBF (this standard)0.4TB/s to 3.0TB/s (3 grades)Up to 512GBUCIeEnterprise NAND SSD (PCIe 5.0)Sequential read up to ~13GB/sTens of TB per drive (models over 30TB available)PCIe 5.0 NVMe

Note: HBF is a newly established tier that fills the gap between HBM (speed) and SSD (capacity). HBM3E and enterprise SSD figures are reference values based on Samsung Electronics’ HBM3E and PM1743 product specifications, respectively; HBF figures are based on the standard specification released this time.

This specification is the first outcome of the collaboration between SK Hynix and SanDisk, which began their HBF standardization partnership in August of last year and officially launched the consortium in February of this year—roughly six months ago. Google and AI semiconductor company Tenstorrent are currently participating in the HBF consortium. SK Hynix plans to expand the HBF ecosystem through open collaboration and accelerate its push into the AI storage market.

SK Hynix also showcased its next-generation NAND technology at the event. The company gave the first public showing of its 10th-generation (V10) 375-layer 4D NAND wafer, developed with optimization for AI infrastructure, along with products incorporating the technology. It plans to begin mass production of high-performance, high-capacity enterprise SSDs (eSSDs) based on this technology early next year. The 375-layer 4D NAND delivers 2.5 times better performance per watt compared to the previous generation, optimized for AI infrastructure environments like data centers that demand both power efficiency and performance.

Kim Chon-sung, Head of Solution Development at SK Hynix, said, “As AI adoption rapidly expands, we are at a point where a fundamental redesign of the overall data processing architecture is required. Through HBF, we will expand the boundaries of memory and storage and contribute to implementing new architectures that enhance system-wide efficiency.”

Alper Ilkbahar, Chief Technology Officer (CTO) of SanDisk, commented, “AI inference is creating new memory demands, and HBF technology is designed to meet those demands. This standard provides system designers with a practical path to bring high-capacity, high-bandwidth memory closer to computing.”

At FMS 2026, SK Hynix is also presenting AI infrastructure solutions based on tiered memory, including HBF. On the first day of the event, Kim Chon-sung and Kang Wook-sung, Head of Next-Generation Product Planning, will deliver a joint keynote address, explaining the need for a tiered memory architecture that connects memories with different characteristics into a single framework to address the surge in data processing volumes driven by the spread of agentic AI. On August 6, a panel discussion will be held featuring Lim Eui-cheol, Head of Solution AT at SK Hynix, along with representatives from Google DeepMind and SanDisk. The panel will discuss changes in the memory structure of AI infrastructure and the role of HBF under the theme “Crossing the Memory Wall with HBF.”

Earlier, at the TSMC Technology Symposium in April, SK Hynix presented a roadmap to expand its next-generation AI memory product lineup beyond HBM to include HBF and 3D stacked DRAM. Industry observers expect that as AI infrastructure investment grows, the competition to establish standards for securing the next-generation memory market beyond HBM will intensify.

Meanwhile, SanDisk, according to market research firm TrendForce, is tied with Micron for fourth place in the global NAND flash market, trailing Samsung Electronics, SK Hynix, and Kioxia. In the first quarter of this year, SanDisk’s global NAND flash market share stood at 13.9% by revenue, matching Micron. SanDisk plans to begin HBF sample shipments in the second half of 2026, with a target of AI inference device samples featuring HBF by early 2027.