High Bandwidth Flash Consortium Tackles AI Memory Bottleneck with New Spec

The High Bandwidth Flash (HBF) consortium has released its first technical specification to tackle AI's memory bottleneck. The spec outlines a new class of NAND-based memory designed for AI inference workloads without the high price tag of high-bandwidth memory (HBM). Led by Sandisk and SK Hynix, the consortium aims to build an open standard for flash memory that can support capacities up to 512 GB per package.

The initial HBF specification promises performance targets with bandwidth ranging from hundreds of GB/s in baseline configurations to a peak of 3 TB/s when using UCIe connections. This tiered memory approach proposes using HBM for the hottest data and HBF for model parameters that need fast access but not the absolute fastest.

The consortium isn't operating in a vacuum, as Samsung has been developing its own approaches to high-bandwidth flash memory outside this framework. The decision by Sandisk and SK Hynix to pursue an open specification rather than proprietary solutions is a deliberate strategic choice.