Neuromorphic Computing

Chips that compute like brains: spiking neurons, event-driven processing and in-memory computation — from Carver Mead's 1980s vision to TrueNorth, Loihi, SpiNNaker and event cameras.

Why imitate the brain in hardware?

Conventional computers follow the von Neumann architecture: a processor and a separate memory, with data shuttled back and forth on a clock. This shuttling consumes much of the energy in modern AI. Brains are organised differently: memory (synaptic strengths) and computation (neurons) are intertwined, millions of units run in parallel, and neurons communicate with sparse, asynchronous spikes — sending signals only when needed.

Origins

In the late 1980s the Caltech engineer Carver Mead coined the term neuromorphic for analogue electronic circuits that mimic neural systems, building a "silicon retina" and "silicon cochlea" with Misha Mahowald and others.

Neuromorphic chips

Digital spiking chips have since grown in scale. IBM's TrueNorth (2014) contained a million spiking neurons and 256 million synapses on a single chip, consuming about 70 milliwatts. Intel's Loihi (2017) and Loihi 2 (2021) support on-chip learning; in 2024 Intel announced Hala Point, a system of 1,152 Loihi 2 chips with about 1.15 billion neurons. The SpiNNaker machine at the University of Manchester, designed by Steve Furber (a co-designer of the original ARM processor), uses a million processor cores to simulate spiking networks in real time; SpiNNaker2 extends this. BrainScaleS in Heidelberg uses analogue circuits running faster than biological time.

Event cameras

A practical neuromorphic success is the event camera (dynamic vision sensor). Inspired by the retina, each pixel reports only changes in brightness, independently and asynchronously, rather than full frames. They have microsecond time resolution, high dynamic range and low power, useful for fast robotics, drones and automotive applications.

Challenges

Spiking networks are harder to train than conventional deep networks, software tools are less mature, and for many tasks GPUs remain more practical. Emerging devices such as memristors, whose resistance changes with history like a synapse, may enable denser, analogue in-memory computing. Neuromorphic engineering is also a tool for neuroscience, testing ideas about how biological circuits compute efficiently.