Energy Efficiency in AI Memories Boosted via Nanoscale Heat Control

Seeking solutions to the energy consumption and heat issues of artificial intelligence systems, researchers led by Prof. Tae Geun Kim at Korea University have designed a dual-confined phase-change heterostructure memory device that controls heat at the nanoscale.

Artificial Intelligence and the Energy Problem

Today's artificial intelligence technologies cause significant increases in electric power and heat while processing massive datasets. At the center of this situation is computer memory, which struggles to transfer data without excessive energy expenditure.

Phase-change memory stores information by melting and freezing tiny pockets of material. However, fitting multiple bits of data into the cells is quite difficult due to heat-induced material degradation and data drift.

Dual-Confined Architecture Developed

Prof. Tae Geun Kim and his colleagues from Korea University eliminated this obstacle by shaping heat within a nanometer-thin material sandwich.

Published in the International Journal of Extreme Manufacturing, the research involved engineering a dual-confined structure stacking antimony telluride layers with nickel ditelluride and molybdenum ditelluride.

Technical Achievements and Performance

This layered geometry acts like a two-stage thermal dam, trapping heat in specific regions depending on the voltage pulse.

The developed cell completes multi-level switching transitions in 50 nanoseconds, reduces reset energy demand by up to 76 percent down to the 115 pJ level, and decreases data drift by more than thirtyfold.

Durability and Ease of Manufacturing

The new memory cell demonstrates endurance of over one million write cycles without experiencing physical degradation.

Additionally, the system provides manufacturing ease by utilizing standard sputtering equipment found in modern semiconductor foundries.