As AI systems demand faster, denser memory, researchers in Shanghai have developed a flash device that operates at picosecond speeds — breaking records and potentially opening the door to a new generation of computing.

Developed by a research team at Fudan University, the device — named PoX — achieves 400-picosecond operation, surpassing the speed of today’s fastest volatile memory. According to the research team’s paper published in Nature on Wednesday, the breakthrough marks the fastest semiconductor charge storage ever recorded, with performance on par with cutting-edge computing systems.

If scaled for industrial use, the technology could reshape memory architecture entirely, blurring the line between storage and memory. Future computers may no longer need separate tiers for data access, enabling faster, more efficient local deployment of large AI models.

“In the blink of an eye, the super flash memory performed 1 billion operations, while conventional USB drives can only manage 1,000,” said Professor Zhou Peng, a lead researcher on the project. 

Today’s AI systems require memory that’s both fast and capacious — but current technologies force a compromise. Volatile memory offers speed but loses data when powered off and consumes significant energy, while flash storage is stable and efficient but dramatically slower. This tradeoff has become a key bottleneck for AI development.

To address this challenge, the Shanghai team reengineered storage from the ground up. “Previously, the approach to speeding up flash memory was to have electrons build up speed first, like warming up on a track,” said Liu Chunsen, a researcher on the project.

Their solution uses specialized materials that allow electrons to move directly to storage sites without delay. The result: up to 2.5 billion operations per second — outpacing the fastest memory today, while remaining energy efficient and scalable.

The team aims to reach megabyte-level integration within three to five years, paving the way for industrial applications.

(Header image: VCG)