High Bandwidth Flash (HBF) Specification: SK Hynix and Sandisk FMS 2026 Announcement

• The specification is open and royalty-free to drive broad industry adoption .
• It builds on the HBF consortium workstream that began in February 2026, itself following an MOU signed in August 2025 .

Technical Details

Stack Configurations and Capacity

• NAND stacks: 8-high and 16-high configurations .
• Maximum capacity per module: up to 512 GB .
• First-generation modules (Sandisk) use a 16-high stack .

Bandwidth Grades

The spec defines three performance grades :

Sandisk has stated its first-generation HBF will target read bandwidth up to 1.6 TB/s, roughly equivalent to one HBM4 stack but at approximately 10x the capacity .

Interconnect

• The standard uses a parallel, high-speed memory bus directly connected to the GPU/accelerator, similar to the HBM interface, rather than a PCIe/NVMe storage protocol .
• Physical and electrical specifications are published under the OCP open hardware framework .

Consortium and Industry Partners

The HBF consortium is led by SK hynix and Sandisk. Two major industry players have publicly joined:

• Google — joined to support the ecosystem .
• Tenstorrent — the AI chip startup also joined the alliance during the standardization process .

Additional participants have not been officially named, but Samsung is reported to be developing its own HBF technology (see competitive landscape).

Commercialization Timeline

• HBF memory samples: expected in the second half of 2026 from Sandisk .
• First AI inference devices using HBF: expected to begin sampling in early 2027 .
• Mass production target: SanDisk is planning a pilot production line by H2 2026, with volume production in 2027 .
• The FMS announcement did not update the product schedule; these timelines remain from prior company guidance .

Competitive Landscape

The HBF space is emerging, with three distinct approaches:

Key differences between the approaches

• Architecture gap: HBF (SK hynix/Sandisk and Samsung) connects NAND directly to the GPU memory bus via TSV stacking — achieving 1.6+ TB/s bandwidth, rivaling HBM. Kioxia's approach keeps the NAND on PCIe, yielding ~64 GB/s bandwidth — far lower, but with higher capacity per module (5 TB) and much simpler integration .
• Open vs. proprietary: The SK hynix/Sandisk spec is open via OCP. Samsung and Kioxia are pursuing proprietary designs.
• Timing: The consortium is approximately 12–18 months ahead of Samsung's publicly stated timeline; Kioxia's GP Series SSDs are closer to production but serve a different performance tier.