Neuromorphic Computing Is Ready for the Big Time

Efforts to build brain-inspired computer hardware have been underway for decades, but the field has yet to have its breakout moment. Now, leading researchers say the time is ripe to start building the first large-scale neuromorphic devices that can solve practical problems.

The neural networks that have powered recent progress in artificial intelligence are loosely inspired by the brain, demonstrating the potential of technology that takes its cues biology. But the similarities are only skin deep and the algorithms and hardware behind today’s AI operate in fundamentally different ways to biological neurons.

Neuromorphic engineers hope that by designing technology that more faithfully replicates the way the brain works, we will be able to mimic both its incredible computing power and its energy efficiency. Central to this approach is the use of spiking neural networks, in which computational neurons mimic their biological cousins by communicating using spikes of activity, rather than the numerical values used in conventional neural networks. But despite decades of research and increasing interest from the private sector, most demonstrations remain small scale and the technology has yet to have a commercial breakout.

In a paper published in Nature in January, some of the field’s leading researchers argue this could soon change. Neuromorphic computing has matured from academic prototypes to production-ready devices capable of tackling real-world challenges, they argue, and is now ready to make the leap to large-scale systems. IEEE Spectrum spoke to one of the paper’s authors, Steve Furber, the principal designer of the ARM microprocessor—the technology that now powers most cellphones—and the creator of the SpiNNaker neuromorphic computer architecture.

Steve Furber on…

In the paper you say that neuromorphic computing is at a critical juncture. What do you mean by that?

Steve Furber: We’ve demonstrated that the technology is there to support spiking neural networks at pretty much arbitrary scale and there are useful things that can be done with them. The criticality of the current moment is that we really need some demonstration of a killer app.

The SpiNNaker project started 20 years ago with a focus on contributing to brain science, and neuromorphics is an obvious technology if you want to build models of brain cell function. But over the last 20 years, the focus has moved to engineering applications. And to really take off in the engineering space, we need some demonstrations of neuromorphic advantage.

In parallel over those 20 years, there’s been an explosion in mainstream AI based on a rather different sort of neural network. And that’s been very impressive and obviously had huge impacts, but it’s beginning to hit some serious problems, particularly in the energy requirements of large language models (LLMs). And there’s now an expectation that neuromorphic approaches may have something to contribute, by significantly reducing…