Skip to content

Neuromorphic chips: the AI that learns to save energy

Neuromorphic processors from Intel and IBM promise to run local AI with a fraction of current energy consumption, bringing intelligence to sensors and batteries.

3 min read
Neuromorphic chips: the AI that learns to save energy

Chips that mimic the brain

Pumping data to data centers may be on borrowed time, at least according to the neuromorphic chip industry: processors designed to work like biological neural networks, processing information in bursts and barely wasting energy. Unlike GPUs, which execute massive calculations in parallel through brute force, these circuits only "fire" when there's something to compute, just like a real neuron does.

The stakes are high: it's estimated that training and operating large AI models already consumes about 1.5% of global electricity on its own, and the figure roughly doubles every two years. If AI wants to reach sensors, glasses, vehicles, or satellites, it can't drag a data center behind it.

The key players: Loihi, NorthPole, and beyond

Two projects draw most of the attention. The first is Loihi 2, Intel's neuromorphic processor, which integrates about eight million digital "neurons" arranged in an asynchronous architecture: each core only activates when it receives current spikes. The second is NorthPole, from IBM, a brain-inspired system that places memory alongside computation, eliminating the classic bottleneck of the von Neumann architecture (the famous "memory wall").

Behind them come players like BrainChip, with its Akida chip, and universities exploring memristors, components capable of remembering how much current passed through them and emulating a synapse with a single analog device.

Why energy consumption matters

The numbers coming out of laboratories are striking:

  • Neuromorphic prototypes consume up to 100 times less in pattern detection tasks than an equivalent GPU.
  • Sensors connected to these chips can operate for months on a button cell battery, compared to hours or days in traditional designs.
  • They respond in microseconds, making them ideal for autonomous cars, robots, and prosthetics.

Real-world applications, not science fiction

The leap from lab to market is underway. In energy, networks of neuromorphic sensors monitor power lines without needing recharged batteries. In healthcare, labs are testing cochlear implants and artificial neurons to restore vision or movement. And in industrial monitoring, these chips detect defects on assembly lines working at the edge, without leaving the factory.

The goal isn't to compete with generative AI, but to bring it where it currently can't reach: devices that run on a single battery.

The challenge of programming neurons

There are still obstacles. Programming a chip that learns through Hebbian reinforcement ("neurons that fire together, wire together") requires new tools, far from the mature ecosystem of frameworks like PyTorch or TensorFlow. Moreover, the industry lacks standards for describing spiking neural networks, although initiatives like snnTorch or Intel's Lava toolkit are accelerating the adoption curve.

The transition is reminiscent of the shift from CPU to GPU for deep learning in the early 2010s. Back then, no one knew how to properly program a GPU; today, they dominate the AI market. Neuromorphic computing could be the next generational shift.

Sources

Intel Research — Neuromorphic Computing
IBM Research — NorthPole
Nature — Neuromorphic Engineering
IEEE Spectrum — Spiking Neural Network Computing

  • #chips neuromórficos
  • #Inteligencia Artificial
  • #hardware
  • #eficiencia energética

All news