SK hynix Study Shows 3D FeNAND Could Turn NAND Into AI Compute Memory

Why AI needs a different kind of NAND

The work is presented in the second part of the SK hynix Newsroom series [Inside the Research]. The paper, titled "Co-Optimizing Cell, Non-Cell, and Page Schemes for Energy Efficient Analog Computing in 3D FeNAND", was published at the 2025 IEEE International Electron Devices Meeting (IEDM). The authors are Won-Tae Koo and Jihun Kim, Technical Leaders at Global RTC, SK hynix.

The starting point is a familiar problem. As large-scale AI models grow, the power consumption and latency of moving data are becoming major bottlenecks in AI infrastructure. The team focused on analog compute-in-memory (A-CiM), an architecture that combines data storage and computation. Instead of continuously transferring data to compute devices, it processes data within or close to memory.

A-CiM uses the electrical states stored in memory cells to perform multiply-and-accumulate (MAC) operations, a fundamental building block of AI computation. Because model weights must be stored and repeatedly accessed, the architecture needs to hold large amounts of data at high density.

Why ferroelectric NAND

Commercial 3D NAND mainly uses a charge trap nitride (CTN) structure, which controls cell states by trapping or releasing electrons in a nitride layer. SK hynix says this offers density advantages but makes it hard to precisely distinguish the multiple current levels analog computing requires.

3D FeNAND instead uses ferroelectric materials, which keep their internal polarization state after an external electric field is removed. Changing that polarization stores data and controls the state of a cell. According to the company, this gives FeNAND an advantage over conventional 3D CTN NAND in precisely controlling multiple current levels at lower power.

The researchers stress that changing the cell structure alone is not enough. Significant power can be lost in the interconnects, such as word lines, that select and control cells. And if pages are not used in line with the AI model's data size and computation sequence, unnecessary switching and latency can result.

Three areas optimized together

The team split the bottlenecks into three areas:

• Cell scheme: how data is represented through different current states.
• Non-cell scheme: the interconnects and peripheral structures used to select and control cells.
• Page scheme: how data used in AI computation is organized into units and sequences.

SK hynix says a bottleneck in any one of these limits overall performance. Precise cells do not help if word line switching overhead is high, and lower interconnect overhead does not help if data organization within a page raises latency. Applying co-optimization across all three areas improved the computational throughput and energy efficiency of 3D FeNAND-based A-CiM, the company reports.

The company attributes the gains to the high integration density of 3D FeNAND and the reduced data movement achieved through the co-optimization. It concludes that improvements beyond what higher memory density alone can deliver can come from system-level optimization.

What it means for NAND's role

SK hynix frames the research as expanding the potential role of NAND in AI systems. In its words, NAND is evolving from storing large volumes of data toward supporting high-density storage of weights for large-scale AI models, reducing the energy burden of data movement and rewriting, and improving computational efficiency.

The company says the high integration density, low power consumption and multi-level control of 3D FeNAND-based A-CiM show the potential for NAND to expand from a storage medium into a computing medium. The source presents this as research potential, and it does not give a product timeline or availability.