SanDisk & SK hynix unveil HBF spec: 512GB GPU memory at 3 TB/s bandwidth via UCIe

SanDisk and SK hynix formally introduced the High Bandwidth Flash (HBF) specification through the Open Compute Project (OCP), positioning it as an open standard for pairing NAND flash non-volatility with HBM-class performance for AI inference workloads. HBF packages support capacities up to 512GB using specialized 8-Hi or 16-Hi NAND die stacks with performance divided into three bandwidth tiers ranging from approximately 0.4 TB/s to 3.0 TB/s. The peak implementation (3.0 TB/s) exceeds a single HBM4 stack (2.0 TB/s) in bandwidth, though HBM4 retains latency advantages.

The spec adopts the Universal Chiplet Interconnect Express (UCIe) standard for heterogeneous integration, simplifying adoption across AI accelerator platforms. SanDisk expects multi-generation roadmap with escalating implementations. To achieve 400 GB/s from a single 512GB package, engineers designed it with 16 core dies featuring concurrent access paths via UCIe at up to 64 GT/s and 64 lanes. HBF targets a new memory tier for inference: combining near-memory bandwidth and NAND persistence, ideal for workloads requiring much larger memory pools than HBM4's 64GB maximum can economically provide. However, uptake is uncertain—despite two years of collaboration between SanDisk and SK hynix, only Google and Tenstorrent have publicly expressed interest in the HBF consortium. Chipmakers including AMD, Broadcom, Intel, Nvidia, Marvell, Micron, Qualcomm, Samsung, and Western Digital have not committed.

For practitioners, HBF signals an alternative memory path for long-context inference and retrieval-augmented generation if ecosystem support builds. The open standard positioning and UCIe compliance suggest modularity, but the lack of major GPU/accelerator vendor commitment creates uncertainty around actual adoption timelines. Anyone designing inference clusters should monitor this specification's evolution but not assume availability in near-term designs.