SK hynix, SanDisk publish first High Bandwidth Flash spec; NAND tier between HBM and SSDs for AI inference

SK hynix and SanDisk released the first Open Compute Project (OCP) technical specification for High Bandwidth Flash (HBF) on August 3–4, 2026, just six months after the HBF workstream launched in February. HBF is a new memory tier positioned between High Bandwidth Memory (HBM) and SSDs, designed to provide high-bandwidth NAND capacity for AI inference workloads. Google and Tenstorrent participated as consortium members in the standardization. The specification defines capacities up to 512GB using 8-high and 16-high NAND die stacks, with three bandwidth grades delivering performance from approximately 0.4TB/s to 3.0TB/s, and adopts the industry-standard UCIe 1 chiplet interface for flexible processor integration.

The HBF specification outlines electrical characteristics, reliability guidelines, packaging standards, and software I/O protocols for die-stack processes. SanDisk previously simulated Llama 3.1 405B inference performance with HBF, reporting results within 2.2% of unlimited HBM capacity while offering 8–16x more capacity at similar cost per die. SanDisk targets Q4 2026 for first HBF samples, with AI inference devices sampling in early 2027. SK hynix also showcased its V10 tenth-generation 375-layer 4D NAND at the Future of Memory and Storage conference, claiming 2.5x performance-per-watt improvement over the prior generation—targeting enterprise SSDs for 2027 mass production.

For AI infrastructure architects: HBF standardization resolves a critical gap in memory hierarchy. Frontier-model inference (GPT-4 Turbo, Claude 3.5, Llama 405B) requires both low-latency access to active tokens and vast storage for model weights. HBM is expensive and power-constrained; conventional SSDs are too slow. HBF bridges that gap at ~1/8 the cost per GB while maintaining near-HBM bandwidth. The open spec via OCP signals a market-wide shift toward modular, reusable memory topologies—expect 2027 designs to standardize HBM for hot working sets and HBF for weight storage, reducing per-inference TCO significantly.