Spintronics hardware for energy-efficient AI
Artificial intelligence is transforming the way we work and live, but training AI models requires enormous amounts of computing power and energy. Finding new ways to make AI faster and more energy efficient is becoming one of the biggest challenges in the field.
Researchers from the R. Ferreira group at INL have developed a new type of spintronic hardware that allows artificial intelligence to potentially learn directly on the chip, making the learning process more efficient. The new technology is based on nanoscale spintronic devices called magnetic tunnel junctions, which are promising building blocks for brain-inspired computing. These devices can process information in complex, neuron-like ways (non-linearity) while retaining information even when power is removed (non-volatility), two key properties for designing energy-efficient AI hardware.
Instead of relying on simplified digital models to train the system, INL researchers Catarina Pereira, Alex Jenkins, Ensieh Iranmehr, Luana Benetti, Subhajit Roy and Ricardo Ferreira, in collaboration with researchers from Politecnico di Bari, Istituto Nazionale di Geofisica e Vulcanologia, University of Porto and University of Messina, developed a new approach that allows the hardware to calculate the information needed for learning directly on the chip.
The study, published in Nature Communications, experimentally demonstrated that this alternative enables spintronic neural networks to learn directly in hardware. Computer simulations further showed that this method can be extended to much larger and deeper neural networks.
Lead author Catarina Pereira further explains: “Using this gradient estimation method, the non-linearity of the devices can be harnessed to perform learning directly in hardware, without relying on digital models of device behaviour.”
The work was supported by the European Union-funded projects RadioSpin, SWAN-on-chip, and NIMFEIA.
Reference Trainable neuromorphic spintronic hardware Via analog finite-difference gradient methods
Catarina Pereira, Alex Jenkins, Eleonora Raimondo, Mario Carpentieri, Ensieh Iranmehr, Luana Benetti, Subhajit Roy, Ricardo Ferreira, Joao Ventura, Giovanni Finocchio, and Davi Rodrigues
