High-Bandwidth Flash aims to bridge HBM capacity limits, Sandisk shares 1.6 TB/s Gen1 target
SK hynix and Sandisk say they have launched a dedicated workstream under the Open Compute Project to standardize High Bandwidth Flash (HBF). The companies describe HBF as a new memory tier for AI inference that sits between HBM and SSDs, targeting higher capacity than HBM without falling back to storage-style access patterns.
Sandisk published adds more detail on what it thinks this layer looks like. Sandisk positions HBF as NAND-based and claims it can deliver “up to 8x to 16x” the capacity of HBM, with similar bandwidth and similar cost in the intended system design.
© Sandisk For first generation hardware targets, Sandisk claims up to 1.6 TB/s read bandwidth, 256Gb per die, and up to 512GB total capacity per 16-die stack. The company also claims the stack is designed to closely match HBM4 in physical footprint, power profile, and stack height, which implies an integration model similar to today’s HBM stacks on AI accelerators.
© Sandisk Sandisk also claims HBF can deliver performance within 2.2% of “unlimited-capacity HBM” in internal testing and simulation, using an inference workload model based on an 8-bit pretrained Llama 3.1 405B parameter setup. The footnote notes the comparison assumes HBM has unlimited capacity for modeling purposes, so it is not measuring the capacity advantage directly.
On the technology side, Sandisk ties HBF to its BiCS NAND roadmap and its CBA (CMOS directly Bonded to Array) approach, plus proprietary 3D stacking aimed at 16-die stacks with reduced warpage and better thermal conductivity. The fact sheet also highlights a non-volatile characteristic and says HBF does not require refresh power, unlike DRAM-based memory.
HBF Gen2 and Gen3
© Sandisk Sandisk’s roadmap section lists Gen2 and Gen3 “preliminary read bandwidth” targets of over 2 TB/s and 3.2 TB/s, with stack capacities up to 1TB and 1.5TB, plus lower power targets versus Gen1 at 0.8x and 0.64x. SK hynix and Sandisk have not published an implementation timeline in this material, but they have already framed standardization as the next step via OCP.
Source: SK Hynix PR
