University of Minnesota CRAM Breakthrough: New AI Hardware Reduces Energy Consumption by 1,000x

University of Minnesota Twin Cities engineers achieved a revolutionary breakthrough in AI hardware efficiency on July 26, 2024, demonstrating a computational random-access memory (CRAM) device that could reduce energy consumption for artificial intelligence applications by a factor of at least 1,000.

Published in npj Unconventional Computing, the research represents over two decades of development, creating the first experimental demonstration of CRAM where data never leaves the memory array for processing.

"This work is the first experimental demonstration of CRAM, where the data can be processed entirely within the memory array without the need to leave the grid where a computer stores information," said Yang Lv, postdoctoral researcher and first author.

The International Energy Agency forecasts AI energy consumption will double from 460 TWh in 2022 to 1,000 TWh in 2026 - equivalent to Japan's entire electricity consumption. The CRAM-based machine learning accelerator addresses this crisis with dramatic efficiency improvements:

The technology leverages spintronic devices using electron spin rather than electrical charge, providing higher speed, greater resilience, and dramatically improved energy efficiency. One MTJ can perform the same function as four or five traditional transistors at a fraction of the energy cost.

"Our initial concept to use memory cells directly for computing 20 years ago was considered crazy," said Distinguished Professor Jian-Ping Wang, senior author. The interdisciplinary team spanning physics, materials science, computer science, and engineering proved the technology's feasibility.

Currently, the team plans collaboration with semiconductor industry leaders to provide large-scale demonstrations and advance AI functionality through this groundbreaking hardware approach.