SK Hynix says the future of memory lies somewhere between HBM and SSDs

SK Hynix has added to the long list of memory types you can’t buy by announcing the first standard specifications for High Bandwidth Flash (HBF), a new class of memory that sits between High Bandwidth Memory (HBM) and the NAND flash used in conventional SSDs. Developed in collaboration with Sandisk, HBF aims to combine much of HBM's bandwidth with the higher capacities enabled by NAND flash, addressing one of the biggest bottlenecks in modern memory hierarchies.

HBF is intended to function as an intermediate memory tier that offers substantially more capacity while maintaining significantly higher throughput than traditional storage. The newly published specification defines capacities of up to 512 GB using either 8-high or 16-high NAND stacks. Performance is divided into three bandwidth classes ranging from roughly 0.4 TB/s to 3.0 TB/s, allowing system designers to scale performance according to workload requirements.

One of the more notable aspects of the specification is its adoption of the Universal Chiplet Interconnect Express (UCIe) standard. Existing HBM typically relies on vendor-driven, package-specific interfaces between the memory stack and processor, while NAND flash communicates over storage protocols such as PCIe/NVMe or SATA after passing through an SSD controller. By adopting UCIe, HBF gains an open, high-speed chiplet interconnect that should make it easier to integrate with a wider range of future processors and accelerators. The specification also establishes electrical characteristics, packaging and reliability guidelines, along with software I/O recommendations.

SK Hynix has published the specification through the Open Compute Project (OCP), allowing for broader industry adoption. The HBF consortium currently includes Sandisk, Google and Tenstorrent, with SK hynix saying it intends to expand participation as the ecosystem matures. Its emergence as an open industry standard suggests the industry is increasingly looking beyond simply adding more HBM, instead adopting a tiered memory architecture that balances bandwidth, capacity and scalability.

That openness also means HBF isn't exclusive to the consortium. Should the standard gain industry-wide traction, other memory manufacturers, including China's CXMT, could eventually develop their own compatible implementations. Whether HBF eventually finds its way into consumer hardware remains to be seen. For now, the technology is aimed squarely at high-performance and enterprise systems that need to move large amounts of data without paying the full cost of HBM. The standard has only just been published, so commercial implementations are likely still several years away. And who knows? The dreaded memory shortage might even be over by the time a single HBF wafer sees the light of day.