SanDisk and SK hynix have jointly published the HBF (Host-to-Buffer) specification, describing a new memory expansion architecture that can provide GPUs with several terabytes of additional NAND-based capacity using stacked flash modules. The spec details support for up to 16-high NAND stacks and around 3 TB/s bandwidth per device, relying on UCIe interconnect technology. It is designed to offload infrequently accessed data from costly HBM or GDDR memory in high-performance accelerators. The initiative could reduce total system costs for GPU clusters running AI and data analytics workloads, while pushing broader industry adoption of UCIe-based chiplet ecosystems. Vendors are expected to prototype compatible controllers, firmware, and accelerator designs next.
This update represents a notable development in the Ai sector. Organizations and founders tracking this space should evaluate potential strategic and technical implications on their operations.