A new memory technology called SOT-MRAM could dramatically reduce the energy demands of artificial intelligence systems, performing critical computing tasks up to 100 times faster while consuming a fraction of the power required by conventional memory. Researchers at the University of Texas at Austin partnered with Taiwan Semiconductor Manufacturing Company (TSMC) to test this emerging technology on real AI workloads, finding it completes write operations in just 2 nanoseconds while using only 2 picojoules of energy per write.

The timing matters because AI's energy appetite is becoming a genuine bottleneck for the industry. Texas alone is expected to become the U.S. capital of data centers within the next few years, which could increase statewide energy usage by 5 times. As AI companies race to build larger models and deploy them globally, the infrastructure supporting these systems is straining power grids and driving up electricity costs in communities hosting data centers.

What Makes SOT-MRAM Different From Other Memory Technologies?

SOT-MRAM, which stands for spin-orbit torque magnetoresistive random-access memory, uses magnetic properties to store information instead of the electrical charges that conventional memory relies on. The key advantage is that it retains data even when power is turned off, making it inherently more efficient. In testing, the technology completed write operations in 2 nanoseconds, while other memory technologies require between 5 to hundreds of times longer, sometimes taking several milliseconds.

The energy difference is equally striking. SOT-MRAM consumed only 2 picojoules per write operation, whereas competing technologies consume hundreds of picojoules or more for the same task. For context, a picojoule is one trillionth of a joule, so these are tiny amounts of energy, but when multiplied across billions of operations happening in data centers every second, the cumulative savings become substantial.

"The unique combination of speed, energy efficiency and endurance makes SOT-MRAM perfectly suited for AI applications, especially in devices where resources like power and memory are limited," said Sam Liu, first author of the research published in Science Advances and a recent UT Austin Ph.D. graduate.

Sam Liu, First Author and Recent Ph.D. Graduate, University of Texas at Austin

One challenge researchers had to overcome was that SOT-MRAM can only hold two states, 0 or 1, which initially seemed limiting for AI applications that often require more complex representations. The team designed a workaround that leverages this binary nature while maintaining sufficient accuracy for practical AI tasks.

How Could This Technology Reshape AI Infrastructure?

The research team tested SOT-MRAM on several real-world AI tasks to demonstrate its practical viability:

• Neural Network Inference: Running trained AI models to make predictions on new data, the most common use case in production systems.
• Binary Neural Network Training: Teaching AI models using a simplified approach that works well with SOT-MRAM's two-state design.
• Probabilistic Graph Modeling: Processing complex relationships in data, a technique used in recommendation systems and knowledge graphs.

The implications extend beyond just saving electricity. According to Jean Anne Incorvia, the faculty leader on the project, SOT-MRAM could enable AI processing to move from centralized data centers to edge devices, meaning smaller, local AI systems could make decisions without constantly communicating with distant servers.

"We show that SOT-MRAM AI accelerators can provide the energy efficiency, with enough accuracy, to eventually replace CPU-based AI accelerators in edge devices such as sensors," explained Jean Anne Incorvia, associate professor in the Cockrell School of Engineering's Chandra Family Department of Electrical and Computer Engineering.

Jean Anne Incorvia, Associate Professor, University of Texas at Austin

Incorvia offered a practical example: a robotic hand equipped with heat sensors could use local AI to make quick decisions about movement without transmitting signals to a distant server. When extremely high accuracy is needed, the robot could then connect to GPU-based data centers in the cloud. This hybrid approach could reduce both energy consumption and latency, making AI systems more responsive and efficient.

What Challenges Remain Before Widespread Adoption?

While the results are promising, the technology is not yet ready for mass deployment. The research team identified two key areas requiring further refinement. First, they need to optimize the characteristics that enable the chips' speed and efficiency, ensuring the technology performs consistently across different operating conditions. Second, they must reduce variation between individual devices, which currently affects the accuracy of neural networks running on the chips.

The broader context for this research reflects growing urgency around AI's energy footprint. There are fundamentally two ways to reduce AI's environmental impact: make the technology more efficient, or reduce reliance on centralized data centers. SOT-MRAM addresses both strategies simultaneously by enabling faster, lower-power processing and making edge AI more practical.

As data centers continue to expand and AI models grow larger, innovations like SOT-MRAM represent a critical path forward. The technology won't solve the energy challenge alone, but it demonstrates that hardware-level improvements can still deliver dramatic efficiency gains, potentially buying the industry time as it works toward more sustainable power solutions.