Predictive device simulations to enable more energy-efficient AI

AI is everywhere. It’s being integrated into every app, added to every website, and applied to every problem. Behind the shiny new tech is an armada of data centers, and powering those data centers takes a significant chunk of the global power grid.

Mamaluy et al. described an approach to reduce AI energy consumption by designing new devices and interconnects using predictive simulations. These first principles-based simulations could lead to specialized AI device hardware that uses orders-of-magnitude less energy.

After examining the computational costs of generative AI applications, the authors found the vast majority of energy consumption was due to matrix multiplication.

“From a practical point of view, [AI] does matrix-vector multiplications on an extremely large scale,” said author Denis Mamaluy. “What if we somehow were able to accelerate matrix multiplication? What would it mean for AI?”

While AI algorithms or processes can be optimized at the software level, they are ultimately bound by the efficiency of the hardware they run on. But current hardware is only moderately capable of performing matrix multiplication, and potential future hardware — such as neuromorphic, analog, or hybrid architecture — could be significantly better.

Therefore, the authors argue, when designing and optimizing AI software, engineers should include hardware needs as well.

“In the past, optimization concerned mostly higher-level parts, and that allowed researchers to become device agnostic,” said Mamaluy. “The predictive simulation is what we believe was missing because it allows connecting materials and device physics.”

Predictive simulations allow researchers to connect design variables, like geometry, materials, and doping profiles, to circuit metrics like current-voltage curves, capacitances, and interface resistances. These simulations use independent parameter sets to avoid device overfitting and can provide a bridge for engineers to construct hardware capable of meeting their needs.

With this approach, future data centers may be able to supply the computational needs of every AI application without the exorbitant energy use.

Source: “Predictive first-principles simulations for co-designing next-generation energy-efficient AI systems,” by Denis Mamaluy, Md Rahatul Islam Udoy, Juan P. Mendez, Ben Feinberg, Wei Pan, Ahmedullah Aziz, Applied Physics Letters (2026). The article can be accessed at https://doi.org/10.1063/5.0333458 .