Laboratory SOT-MRAM writes in 2 nanoseconds for edge AI

Each write operation took 2 nanoseconds. It costs 2 picojoules of energy. At the University of Texas at Austin, engineers working with Taiwan Semiconductor Manufacturing Company (TSMC) fabricated and tested SOT-MRAM, magnetic memory that keeps its information even when the power is off.

The team put the chips through neural-network inference, binary neural-network training and probabilistic graph modeling. SOT-MRAM uses magnetic properties, and the researchers designed their system around its ability to store only two states—0 and 1—while retaining enough accuracy for those AI tasks. The result is a memory technology the researchers say combines speed, lower energy use and endurance.

That efficiency matters beyond the chip. AI and the data centers supporting it are driving energy demand in Texas and worldwide; Texas is expected to become the U.S. capital of data centers in the next few years, with a possible 5x increase in statewide energy use. More efficient memory could reduce the energy used by AI hardware, while local processing could reduce how much work has to travel to those data centers.

So what changes in practice? Sensors and other edge devices with tight power and memory limits could eventually make quick, sufficiently accurate decisions on their own. The researchers use a heat-sensing robotic hand as an example: local AI could move it without sending the neural signal to the cloud, while a GPU-based data center could handle cases requiring very high accuracy.

The researchers plan to refine the characteristics that give the chips their speed and efficiency and reduce variation between devices, because that variation can reduce neural-network accuracy.