SK hynix and SanDisk Launch HBF Standard: Filling the AI Inference Memory Gap Between HBM and SSDs
HBM is not enough and SSDs are too slow — so SK hynix and SanDisk decided to build a new layer in between.
Three Key Takeaways
First HBF standard specification officially published. At the Flash Memory Summit (FMS) 2026, held August 4–6 in Santa Clara, SK hynix and SanDisk jointly released the first OCP technical specification for High Bandwidth Flash (HBF). The positioning is precise: sitting between HBM and SSDs, HBF uses NAND flash to deliver bandwidth approaching HBM levels while preserving flash's massive capacity advantage. The spec defines three bandwidth grades ranging from roughly 0.4 TB/s to 3.0 TB/s, with capacities up to 512 GB using 8-high and 16-high NAND die stacks.
Targeting AI inference's memory bottleneck. The pain point in AI inference is increasingly clear: models run into the hundreds of billions of parameters, HBM is expensive and capacity-limited, and loading weights purely from SSDs introduces unacceptable latency. HBF aims to provide a sweet spot — enough bandwidth for real-time inference, enough capacity to hold large model weights, and a cost far below equivalent HBM.
An industry consortium is already forming. HBF is not a two-company side project. The consortium already includes Google and Tenstorrent. On August 6, executives from SK hynix, SanDisk, and Google DeepMind are holding a joint panel to discuss HBF's role in overcoming AI memory limitations. SK hynix is also showcasing its tenth-generation 375-layer 4D NAND technology for the first time at FMS.
WangDou's Take
The AI chip industry has an open secret: Nvidia GPU compute doubles year after year, but the memory bandwidth feeding those GPUs has never kept pace. HBM is great, but a single HBM4 package holds only a few dozen gigabytes and the per-unit price keeps climbing. So the problem becomes: you have a Ferrari engine with a five-liter fuel tank. HBF's idea is to give you a fifty-liter auxiliary tank — the fill rate is not as fast as the main tank, but it is enough to finish a full lap without pitting. This is not some disruptive breakthrough; it is packaging and interface engineering on top of existing NAND processes. But engineering optimizations are precisely what can be mass-produced and cost-reduced. SK hynix and SanDisk grabbed the standard-setting seat first; everyone else either joins the consortium or invents their own spec — and history tells us that the "own spec" crowd rarely survives past two product generations.
Source: EE Times Asia, HotHardware, SK hynix Newsroom