- Sandisk and SK Hynix offer a large -band memory fed by Flash to manage the larger AI models
- High bandwidth flash could store much more data than HBM based on DRAM for workloads IA
- Nand non-voltatility energy savings could reshape the AI data center cooling strategies
Sandisk and SK Hynix have signed an agreement to develop memory technology that could change the way AI accelerators manage large -scale data.
Businesses aim to normalize “high bandwidth flash” (HBF), an alternative based on the Nand in traditional memory of the high bandwidth used in AI GPUs.
The concept is based on packaging conceptions similar to HBM while replacing part of the dram battery with flash, exchanging a certain latency for a considerably increased capacity and non-voltility.
AI memory batteries to manage larger models to lower power requests
This approach allows HBF to provide between eight and sixteen times the storage of HBM based on DRAM at almost similar costs.
Nand’s ability to keep data without constant power also provides potential energy savings, an increasingly important factor as IA inference extends in environments with strict power and cooling limits.
For hyperscal operators performing large models, the change could help treat both the thermal and budgetary constraints which already serve the operations of the data center.
This plan is aligned with a research concept entitled “LLM in a flash”, which described the way in which language models could operate more effectively by incorporating SSDs as an additional level, raising pressure on the dram.
HBF essentially incorporates this logic into a single wide -band package, potentially combining the storage scale of the largest SSD with the speed profile necessary for AI workloads.
Sandisk presented its HBF prototype to Flash Memory Summit 2025, using bicking techniques of bics NAND and WAFER.
Sample modules are expected in the second half of 2026, the first AI material using HBF projected for the beginning of 2027.
No specific product partnership has been disclosed, but the position of SK Hynix as a major memory supplier for the main manufacturers of AI chiples, including NVIDIA, could accelerate adoption once the standards are finalized.
This decision also comes while other manufacturers explore similar ideas.
Samsung has announced storage levels supported by Flash and continues to develop HBM4 Dram, while companies like Nvidia remain attached to heavy conceptions of Dram.
In case of success, the Sankisk and SK Hynix collaboration could create heterogeneous memory stacks where DRAM, FLASH and other types of persistent storage coexist.
Via Toms equipment
