ChainCatcher reported that researchers from P Equity Research and SemiAnalysis, including Nick Doyle, discussed high-bandwidth flash, or HBF, during an X Space.
Cost outlook is still unclear
Doyle said it is too early to tell how much HBF can narrow its cost premium relative to high-bandwidth memory, or HBM. Most of the data available today comes from vendors, he said. As one example, Sandisk has claimed that HBF could deliver a cost per bit at roughly one-eighth of HBM.
He said yield and testing should improve as production scales. Still, structural costs tied to TSVs, stacking, and pSLC mode would remain. Endurance is another major unknown. If wear turns out to be worse than expected, total cost would rise.
Not a replacement for HBM
In Doyle’s view, HBF has a narrow application range and is aimed only at AI inference, especially low-batch, long-context mixture-of-experts, or MoE, models. He said it is not a substitute for HBM.
The practical bandwidth target is about 1.6 TB/s, which he described as being in HBM3E territory. That profile is better suited to sequential reads for loading model weights and to capacity needs in local or private enterprise deployments with a limited number of GPUs, rather than very large-scale bandwidth-heavy settings.
Thermals and manufacturing remain open questions
Thermal reliability has not been resolved, according to the discussion. Flash placed next to GPUs in high-temperature environments could degrade faster. Possible mitigations such as UCIe separation and daily refreshes have yet to be validated.
On manufacturing, Sandisk and Kioxia bring 3D NAND experience, while SK Hynix adds HBM-style stacking capability. Even so, whether that can translate into volume production remains unproven.
Memory market is becoming more specialized
The broader takeaway from the discussion was that memory is moving toward a more layered and specialized market. NAND shortages are expected to continue through 2028, and HBF could add another variable to supply and demand.