03 · Neuromorphic computing

Silent until something happens

A third way to build a computer: neurons and synapses instead of instructions, events instead of a clock, memory beside the arithmetic. What it does well, and the price it pays.

The GPU← → to move between chapters

01 / 05Established

One neuron

Neuromorphic hardware starts from a different primitive. Not an instruction and not a multiply, but a small analog-like state that leaks, accumulates and occasionally fires.

Integrate

Each incoming spike nudges the membrane potential up or down by the weight of the synapse it arrived on.

Leak

Between spikes the potential decays towards rest with time constant τ. Inputs must arrive close together in time to add up: the neuron is a coincidence detector.

Fire

When the potential crosses the threshold the neuron emits one spike, resets, and is briefly unresponsive. The output is a train of identical events; the information is in their timing.

Try it

Lower τ to 5 ms: inputs no longer overlap and the neuron falls silent at the same input rate. Then raise inhibition and watch it veto the excitatory inputs.

Key numbers

Model: leaky integrate-and-fire
Neurons in a human brain: ≈ 86 billion
Power of a human brain: ≈ 20 W

The LIF model is a deliberate simplification. Brain figures are standard neuroscience estimates; living neurons are outside the SYNAPSE architecture.