The feat, led by Zhou Peng at the State Key Laboratory of Integrated Chips and Systems, vaults non-volatile memory into a realm once thought reserved for the fastest volatile RAM.

Published in Nature, the work replaces sluggish silicon channels with two-dimensional Dirac graphene. It harnesses ballistic charge transport for what Zhou calls “AI-driven process optimisation” to push flash to its theoretical speed limits.

“We drove non-volatile memory to its theoretical limit, and the design paves the way for future high-speed flash memory.”

Dubbed “PoX,” the device uses two-dimensional super‑injection to pump charge into the storage layer at lightning speed, ditching the injection bottlenecks that have dogged flash memory for decades.

Fellow researcher Liu Chunsen said it was like moving from a USB stick that writes a thousand times a second to a chip that fires a billion times in the blink of an eye. For perspective, the prior record for flash programming hovered around two million ops per second — PoX leaves it in the dust.

More importantly, because it’s non-volatile, PoX doesn’t need standby power to retain data, making it ideal for edge AI and battery-starved environments. In AI accelerators, memory bandwidth—not maths—is the biggest energy hog, and this design could solve that problem.

Reviewers call it a “completely original mechanism” that could upend global flash strategies. If PoX scales, it might eliminate the need for SRAM caches in AI chips, enabling low-power, always-on devices and stuffing entire databases into persistent RAM.

The team’s next steps include scaling cell architecture and building full arrays. While they haven’t named any industrial partners, Chinese foundries are already trying to bolt 2D materials onto standard CMOS — and PoX could be the wedge they’ve been waiting for.

Zhou added: “Our breakthrough can reshape storage technology, drive industrial upgrades and open new application scenarios.”