SK hynix has introduced the industry's first official specification for High Bandwidth Flash (HBF), a next-generation memory technology designed for the AI era, in collaboration with SanDisk. The new standard represents a significant step toward building more efficient AI infrastructure by bridging the performance gap between high-bandwidth memory (HBM) and solid-state drives (SSDs).
The specification was unveiled at FMS (Future of Memory and Storage) 2026, which opened on August 4 in Santa Clara, California. It marks the first major outcome of the HBF standardization effort launched by SK hynix and SanDisk last year, followed by the establishment of the HBF Consortium earlier this year.
HBF is designed as a new memory tier positioned between HBM, which handles high-speed AI computation, and NAND flash-based SSDs, which provide large-capacity storage. By combining the high bandwidth of HBM with the scalability and storage density of NAND flash, HBF aims to improve the efficiency of AI systems processing increasingly massive datasets generated by generative AI and agentic AI applications.
The newly announced specification supports memory capacities of up to 512GB and bandwidth ranging from 0.4TB/s to 3.0TB/s. To maximize interoperability across different processors, including GPUs and CPUs, the standard adopts the Universal Chiplet Interconnect Express (UCIe), an open industry interface that enables flexible system integration.
The specification also defines key technical requirements for commercialization, including interface protocols, electrical characteristics, stacking reliability, packaging standards, and software guidelines for data input and output.
Rather than serving as a proprietary technology, the HBF specification has been proposed through the Open Compute Project (OCP), allowing broad industry adoption. The open-standard approach is expected to encourage ecosystem expansion and accelerate the development of next-generation AI memory platforms.
The HBF Consortium currently includes SanDisk, Google, AI processor developer Tenstorrent, and other industry partners. SK hynix said it will continue expanding collaboration with global technology companies to strengthen the HBF ecosystem.
During FMS 2026, SK hynix is also outlining its vision for a "Tiered Memory" architecture. In a keynote presentation, Kim Cheon-seong, Head of Solution Development, and Kang Wook-sung, Head of Next-Generation Product Planning, explain that future AI systems will require multiple layers of memory working together efficiently, as a single memory technology can no longer meet the rapidly growing demands of AI workloads.
Later in the event, representatives from Google DeepMind, SanDisk, and SK hynix will discuss how HBF can help overcome the growing "Memory Wall" challenge, one of the key performance bottlenecks facing large-scale AI infrastructure.
In addition to HBF, SK hynix is showcasing its next-generation NAND technology, including a 375-layer V10 4D NAND wafer and related products. The company said the new generation delivers approximately 2.5 times better power efficiency than its predecessor and will be incorporated into high-performance enterprise SSDs for AI data centers beginning early next year.
The exhibition also features memory solutions for mobile devices, PCs, automotive applications, robotics, and data centers, highlighting SK hynix's strategy to build a comprehensive memory portfolio for the AI era.
Industry observers believe competition for post-HBM memory technologies is accelerating as AI models continue to grow in complexity and data requirements. Technologies such as HBF are increasingly viewed as essential components of future AI server architectures, helping reduce data transfer bottlenecks between computing and storage while improving overall system efficiency.
"HBF represents a critical step toward a new generation of AI memory architecture," Kim Cheon-seong said. "As AI adoption expands, innovation across both memory and storage will become increasingly important, and we intend to lead the development of more efficient AI infrastructure through HBF."
