The hardware behind the future


Samsung Pulls Nvidia Into Its Ferroelectric NAND Lab, and 60 Hours Becomes 10 Seconds

Nvidia researchers are co-authors on a Samsung and Georgia Tech preprint that replaces a slow device simulation with an AI surrogate model. The paper claims a speedup of more than 10,000 times, on one cell configuration.

Samsung Pulls Nvidia Into Its Ferroelectric NAND Lab, and 60 Hours Becomes 10 Seconds

Nvidia buys more memory than almost anyone alive, so when its researchers turn up as co-authors on a paper about a NAND flash cell that does not exist as a product yet, the sensible question is what they are doing there. Samsung Electronics has been chasing ferroelectric NAND for years. Nvidia has now joined that work, with Georgia Tech alongside it.

Ferroelectric materials hold a polarized state without a continuous supply of electricity, which turns out to matter enormously if you are stacking memory cells hundreds of layers high. Conventional NAND uses silicon and needs a firm shove of voltage to move charge around. Put a ferroelectric layer in its place and the shove gets smaller. The research team frames ferroelectric field-effect transistor vertical NAND, shortened to Fe-VNAND, as a way past the limits on scaling cells upward, and it names lower programming voltages as the reason.

Retention is the catch. That is the unglamorous question of whether a cell still remembers what was written to it a year later, and in a ferroelectric cell two effects work against each other at once. Charge leaks out of traps, and the ferroelectric polarization decays on its own. Untangling that interaction is what Technology Computer-Aided Design software exists for, and it is slow enough that sweeping a wide range of candidate designs is not affordable.

So the joint team went after the simulator instead of the cell. Their answer is a Physics-Informed Neural Operator, a model with the governing physics built into its structure rather than learned from examples, trained to predict threshold voltage shift and retention behavior. The framework, the paper states, "achieves a speedup exceeding 10000x compared to TCAD while maintaining physical accuracy".

Seoul Economic Daily, which reported the collaboration, attaches wall-clock numbers to that claim: a run that took roughly 60 hours now lands in under 10 seconds.

Then read the last line of the abstract before ordering champagne. "This study demonstrates the model's effectiveness on a single FeFET configuration," the authors write, describing the work as a pathway toward modeling retention loss. The preprint runs to four pages and is marked as a conference submission. One cell geometry is a starting point, not yet a design tool.

Why Nvidia is in the room at all is the more interesting half. Memory has stopped being a component bought late in the design cycle and become the thing that decides what an AI system can do, and Seoul Economic Daily reports that Samsung intends to work with Nvidia on ferroelectric NAND all the way to commercialization. The author list places Nvidia staff in Santa Clara next to Samsung's Semiconductor Research and Development group and Georgia Tech's school of electrical and computer engineering, which describes a research collaboration rather than a supply contract.

That same report puts Samsung's production NAND stacking in the 200 to 300 layer range today, with ferroelectric materials treated as the route to 1,000. Those figures come from one outlet citing unnamed industry sources, and Samsung has not published them itself. What is on the record is a four-page preprint showing that one very slow simulation can be made very fast, which is the kind of result that decides whether a memory technology gets a decade of engineering behind it or a polite paragraph in somebody's review paper.

Leave a Reply