HBF: a new tier between HBM and SSD
SK hynix and Sandisk have unveiled the first open standard specification for High Bandwidth Flash (HBF), targeting next‑generation AI infrastructure. Announced at FMS 2026 in Santa Clara, the spec formalizes HBF as a new memory tier that bridges high‑bandwidth memory (HBM) and solid‑state drives (SSDs).
HBF combines NAND flash capacity with data‑transfer performance approaching HBM, aiming to relieve bandwidth and capacity bottlenecks in data‑hungry AI inference workloads.
Key points of the initial spec:
- Capacities up to 512 GB per device
- Support for 8‑high and 16‑high stacked NAND dies
- Three bandwidth grades from roughly 0.4 TB/s up to 3.0 TB/s
- Defined electrical characteristics, packaging, reliability, and software I/O guidelines
HBF connects to processors via UCIe (Universal Chiplet Interconnect Express), enabling chiplet‑based integration alongside diverse CPUs, GPUs, and AI accelerators. The spec is published through the Open Compute Project as an industry‑wide open standard, not a proprietary interface. Members of the HBF consortium already include Google and AI processor designer Tenstorrent.
Tiered memory for agentic AI and 375‑layer 4D NAND
At FMS, SK hynix executives presented a Tiered Memory architecture that coordinates multiple memory types within a single system. The company argues that no single technology can simultaneously satisfy the speed, capacity, and efficiency demands of emerging agentic AI applications, making a layered memory hierarchy essential.
A panel with SK hynix, Sandisk and Google DeepMind representatives is set to explore how HBF could reshape AI memory architectures.
SK hynix also showcased its 10th‑generation 375‑layer 4D NAND (V10) technology for the first time. Still under development, the device is claimed to deliver 2.5× better performance per watt than the previous generation, pointing to more energy‑efficient flash for future HBF modules and data‑center storage.
Source: eeNews Europe
