Neuromorphic computing
Chips that work like networks of spiking neurons promise AI at a fraction of the energy.
Computing and AI · New paradigms · No. 475
- Emerging
- Readiness 5 of 9
- Component
A chip with a million artificial neurons ran on 63 thousandths of a watt, less than a small torch bulb.
Chart: Illustration: Neuromorphic computing
Where it comes from
Chart: Map of where it comes from
In 1989 Carver Mead of Caltech proposed chips that work like nerve tissue, building on the 1952 Hodgkin and Huxley model of nerve impulses. Large digital neuromorphic chips followed from IBM, Intel and Manchester.
Three things to know
- Carver Mead, who in 1969 co-proposed very large scale integrated chips, published Analog VLSI and Neural Systems in 1989.
- Intel's Hala Point of 2024 links 1,152 Loihi 2 chips into 1.15 billion neurons and draws at most 2,600 watts.
- Sandia fitted a whole fruit-fly brain map onto 12 Loihi 2 chips and ran it more than 100 times faster than ordinary simulation.
The journey
- 1 Neuron modelBiology shows how nerve cells add up inputs and fire spikes
- 2 CircuitEach artificial neuron and its synapses are built into a small core on a chip
- 3 ChipThousands of cores are made in an ordinary chip factory
- 4 SystemHundreds of chips are linked so spikes can travel between them
- 5 TaskA spiking network is loaded to sense, recognise or simulate in real time