neuromorphic.foundation
neuromorphic computing

Systems

The field is defined as much by the machines built as by theory. These are the systems most often cited; all figures are as published by their builders.

SystemWho and whenWhat is distinctive
Silicon retinaCarver Mead and Misha Mahowald, Caltech; first published 1988, widely known from 1991Among the first neuromorphic sensors: analog circuits that mimic the first layers of the eye
Dynamic vision sensorLichtsteiner, Posch and Delbruck, Zurich; first shown 2006, full paper 2008A camera whose pixels report only changes in brightness, each independently, with microsecond timing
NeurogridStanford University, 2014A board of sixteen mixed analog–digital chips simulating a million neurons in real time
TrueNorthIBM, 2014One million digital spiking neurons and 256 million synapses on one chip, at about 70 milliwatts
SpiNNakerUniversity of Manchester; million-core machine switched on in 2018Very many small conventional processor cores joined by a network built for spikes
BrainScaleSHeidelberg UniversityAnalog neuron circuits that run far faster than biological real time
Loihi and Loihi 2Intel, 2017 and 2021Asynchronous digital chips with programmable on-chip learning
TianjicTsinghua University, 2019A hybrid chip running spiking and conventional neural networks together
Hala PointIntel, installed at Sandia National Laboratories, 2024A system of 1.15 billion neurons, the largest announced at the time
Memristive crossbar arraysMany laboratories, since 2008Grids of resistive memory elements that multiply and sum in place, in the analog domain

Not the same as a neural network accelerator

The chips that train and run today’s large neural networks are fast, dense, clocked arithmetic engines; they borrow the brain’s wiring diagram loosely and none of its signalling. Neuromorphic hardware borrows the signalling: spikes, events, locality. The two fields share a vocabulary and an ancestry in the 1943 neuron model of McCulloch and Pitts, and little else.