Rethinking Computing: Researchers Tackle AI’s Energy Demands

As artificial intelligence (AI) advances, so too does the need for vast computing power—raising urgent questions about energy consumption. At USC’s School of Advanced Computing and Viterbi School of Engineering, researchers are exploring ways to make computing for AI more energy-efficient. Strategies include trimming and simplifying neural networks, transferring learning from pre-trained models, and reducing the bit-depth of data used in machine learning processes. These methods can dramatically shorten training times while lowering hardware requirements and computational costs, according to Professor Massoud Pedram.

Hardware innovation is another key frontier. While GPUs dominate AI processing, their high cost and energy demands have led researchers like Professor Murali Annavaram to suggest pairing them with CPUs to improve efficiency. Other labs are customizing chips to better match the needs of specific AI tasks, optimizing how data is processed and stored. A more radical approach comes from Professor Peter Beerel, whose lab is pioneering superconducting electronics using niobium to eliminate resistance and enable ultra-low-power computing, supported by new algorithms and design automation tools.

Looking to the future, USC researchers are also pushing the boundaries with neuromorphic computing — systems modeled on the human brain. Professors like Joshua Yang and Barath Raghavan are developing technologies like analog in-memory chips and “information batteries” that store energy as pre-computed data. These innovations promise to dramatically reduce energy use while maintaining or even enhancing performance, signaling a transformative shift in how computing systems are designed to meet the growing demands of AI, edge devices, and real-world applications.