Bandwidth vs. IOPS vs. efficiency: Samsung develops three new types of NAND storage for AI applications

This type of memory is specifically designed for local AI applications. To shorten data paths and improve speed and energy efficiency, the NAND chips are stacked in four or eight layers and connected via TSV (through-silicon via) connections. In this case, performance is somewhat less important than efficiency, as AI applications are intended to run locally while consuming as little power as possible.

This variant is aimed more at high-performance AI data centers, so the flash memory is designed to handle particularly large amounts of data per unit of time. One key term here is High Bandwidth Flash (HBF). Ultimately, the goal is to transfer large volumes of data as quickly as possible.

This variant focuses on particularly high IOPS figures, or input/output operations per second. This is especially important for AI inference, where large numbers of small data accesses need to be processed quickly.

Only zNAND-O has been publicly unveiled so far. The other two systems have not been officially confirmed and are still in the early stages of development.

With this move, Samsung is following the strategy of its competitor SK Hynix, which is also pursuing three specialized types of NAND. In the long term, this approach could intensify competition in the AI storage market and expand the role of NAND flash beyond traditional SSD applications.