Hands on
A population of leaky integrate-and-fire neurons running live in your browser. Move a slider and everything responds instantly — the raster plot, the firing rate, and the energy each architecture would have spent.
How many independent cells share the same input. More neurons means more parallel work — and on classical hardware, proportionally more constant power.
The membrane voltage a neuron must reach before it fires. Raise it and the network grows selective and quiet; lower it and everything chatters.
How often the outside world pokes each neuron, in events per second. Think of it as how busy the sensor scene is.
How much voltage a single incoming event delivers. This is the number that learning rules like STDP adjust.
How quickly charge drains away between inputs. Leak is why timing matters: inputs must arrive close together to add up.
Live raster plot
each dot is one spike · vertical = neuron · horizontal = time
0.0 Hz
per neuron
0
0 ms simulated
0 pJ
event-driven
0 pJ
always clocked
collecting…
The conventional bar grows at a constant rate because a clocked processor burns power on every cycle whether or not anything happened. The neuromorphic bar grows only when neurons spike. Turn the threshold up or the input frequency down and watch the gap widen — that sparsity is the whole advantage.
Every horizontal band is one neuron and time flows left to right. A dot appears the instant that neuron fires. Dense vertical stripes mean the population is firing together; scattered dots mean each cell is responding to its own input.
This is a leaky integrate-and-fire model. Charge arrives, the membrane voltage climbs, and a little leaks away each millisecond. When the voltage crosses the threshold the neuron emits a spike, resets to zero and ignores input for a 3 ms refractory period — exactly like a biological cell.
Set the threshold to 2.0 and the weight to 0.1. Almost nothing fires, and neuromorphic energy flatlines while the conventional bar keeps climbing. Now raise the weight to 1.0: the raster fills in and the efficiency advantage shrinks. Sparse inputs are where brain-inspired silicon wins.
Neuromorphic cost is modelled as roughly 25 picojoules per synaptic event plus a tiny leakage term. Conventional cost is a fixed per-neuron-per-millisecond figure, because the clock never stops. These are illustrative orders of magnitude drawn from published Loihi and TrueNorth figures, not measurements.
Guided tour
Each one takes about thirty seconds and teaches a real principle of neural computation.
Keep the weight at 0.35 and slowly raise the spike threshold from 0.2 to 2.5. Around 1.5 the raster empties out: individual inputs can no longer overcome the leak before the charge drains away. This is how a neuron becomes selective — it only responds when enough evidence arrives at once.
Set the threshold to 1.0 and the weight to 0.15, then sweep the input frequency from 5 Hz to 200 Hz. At low rates almost nothing fires, because each small input leaks away before the next arrives. At high rates the inputs stack and the cell fires steadily. Timing, not just magnitude, decides the outcome.
Push the weight to 1.2 and the frequency to 200 Hz. The firing rate does not rise indefinitely — it saturates near 250 Hz, because after each spike the neuron enforces a 3 ms silence. Real neurons have exactly this speed limit, and it stops runaway activity.
Set 120 neurons, threshold 2.0, weight 0.1. Neuromorphic energy barely moves while the conventional bar races ahead, because a clocked processor pays for 120 idle neurons every millisecond. Now drop the threshold to 0.3 and the gap narrows sharply. Efficiency comes from sparsity, not from magic.
Inputs arrive as a Poisson process, which is how sensory neurons are usually modelled: events are random but occur at a set average rate. The energy figures use published orders of magnitude — a synaptic event on neuromorphic silicon costs picojoules, while a clocked core pays a fixed cost every cycle regardless of activity. They are teaching numbers, not benchmark results.