Brain-inspired chips could cut artificial intelligence’s soaring energy needs

With the power consumption of artificial intelligence likely to double by 2026, an international team of researchers has unveiled an ambitious roadmap for brain-inspired computing technology that could significantly reduce the environmental footprint of artificial intelligence while increasing its capabilities.

This comprehensive review published January 22 in the journal Nature by 23 leading experts from academia and industry outlines how neuromorphic computing – chips that mimic the architecture of the brain – could revolutionize everything from smartphones to Everything for a smart city while using only a fraction of the energy of traditional systems.

Gert Cauwenberghs, Distinguished Professor in the Department of Bioengineering at UC San Diego and one of the paper’s co-authors, explained: “Neuromorphic computing is particularly important today as we witness unsustainable expansion of power-hungry and resource-hungry artificial intelligence systems. .

The field appears to be at a critical juncture. “We now face a tremendous opportunity to build new architectures and open frameworks that can be deployed in commercial applications,” said Dhireesha Kudithipudi, Robert F. McDermott Chair at the University of Texas at San Antonio and author of the paper. Corresponding author.

Potential applications range from scientific computing and artificial intelligence to augmented reality, wearables, smart agriculture and urban infrastructure. Recent breakthroughs have demonstrated the technology’s promise—in 2022, a team led by Cauwenberghs developed a chip that can run a variety of artificial intelligence applications using far less energy than traditional systems while maintaining equivalent accuracy .

But researchers believe that to achieve widespread adoption, neuromorphic systems need to better mimic one of the brain’s key features: selective pruning of neural connections. The human brain initially forms a large number of connections and then strategically eliminates most of them, a process that optimizes space efficiency and information retention.

“Scalable scalability and remarkable efficiency result from massive parallelism and hierarchies in neural representations,” Cauwenberghs noted. He described how the system combines dense local connections within core units, such as the brain’s gray matter, with sparse Long distance communication (similar to white matter) combines.

The impact could be huge. At the San Diego Supercomputing Center, which regularly evaluates new computing architectures, researchers see huge potential. “This paper demonstrates the huge potential of neuromorphic computing for large-scale, real-life applications,” said Amitava Majumdar, director of data-enabled scientific computing at the center and co-author of the paper.

The roadmap emphasizes that success will require broad collaboration between academia and industry, as well as the development of more user-friendly programming tools to make the technology accessible to a wider range of developers and researchers.

Rather than proposing a single solution, the authors envision a range of neuromorphic hardware designs tailored for different applications. This flexible approach can help accelerate adoption in a variety of sectors, from healthcare and robotics to environmental monitoring and autonomous systems.

Progress is already underway. Last year, the Cauwenberghs and Kudithipudi received $4 million from the National Science Foundation to launch THOR: The Neuromorphic Commons, a groundbreaking research network that will provide open access to neuromorphic computing hardware and tools to facilitate this Collaborative innovation in rapidly growing fields.