SK Hynix and SanDisk jointly announce HBF (High-Bandwidth Flash) industry standard with Google and Tenstorrent support. OCP technical specification released for standardized high-bandwidth flash architecture.
SK Hynix and SanDisk have unveiled the first standard specification for high-bandwidth NAND flash (HBF), a next-generation storage technology designed for AI accelerators. The announcement was made at FMS 2026, the world's largest memory technology conference, held on August 4th at the Santa Clara Convention Center in California.
HBF represents a fundamentally different approach to AI memory architecture compared to HBM (high-bandwidth memory). While HBM is DRAM-based and operates as a "workbench" for rapid data transfer near AI chips, HBF is a NAND flash-based solution that functions as a "large storage cabinet" adjacent to processors. Because NAND flash can store significantly more data than DRAM in the same physical area, HBF enables faster data exchange with graphics processing units (GPUs) while addressing the bandwidth and capacity scalability challenges that arise from rapidly expanding AI inference workloads.
SK Hynix began standardizing HBF with SanDisk in August 2025 and established a broader consortium alongside Google and Tenstorrent in February 2026. The release of the standard specification just six months after the consortium's formation demonstrates the company's strategy to establish market leadership in HBF technology—a position it previously secured in HBM through collaboration with AMD in 2013.
The HBF standard supports two configurations, with NAND flash dies stacked in either 8 or 16 layers. The specification accommodates a maximum storage capacity of 512GB and defines bandwidth across three performance tiers, ranging from 0.4TB/s to 3.0TB/s. The interface between HBF and GPUs employs UCIe, an industry-standard connection protocol that enables flexible integration across different processor types.
SK Hynix has released the HBF standard as open-source through the Open Compute Project (OCP), the world's largest open data center technology collaboration. This approach facilitates broader industry adoption and standardization, particularly important as competitors including NVIDIA and Kioxia pursue their own proprietary next-generation AI accelerator memory standards.
At FMS 2026, SK Hynix will present its vision for "tiered memory" architecture, which optimizes multiple memory types into unified systems, and will debut its 10th-generation 375-layer 4D NAND flash wafers and products. Kim Cheon-seong, head of SK Hynix's Solution Development Division, stated: "As AI utilization rapidly expands, there is a growing need to redesign the entire data processing structure. Through HBF, we aim to expand the boundary between memory and storage and contribute to implementing a new architecture that enhances overall system efficiency."