Brains and AIMiddleLesson 3005 min read

Why brains use so little energy

The human brain runs on about 20 watts. Learn the main ideas that make brains so efficient: sparse spikes, events instead of clocks, and memory right next to computing.

✦ Ask the tutor 🧪 Lab · 20 min 📝 Lesson test

Speed

Lesson in motion

In 60 seconds

Why brains use so little energy

The human brain runs on about 20 watts. Learn the main ideas that make brains so efficient: sparse spikes, events instead of clocks, and memory right next to computing.

1/4

In simple words

Imagine a school where every student must shout their answer every second, even when they have nothing new to say. It would be loud and tiring. Now imagine a school where students speak only when they have something new. That second school is much quieter and uses far less energy. Brains are like the second school.

The human brain has roughly 86 billion neurons and runs on about 20 watts of power. These are commonly cited estimates. For comparison, 20 watts is in the range of a small LED bulb or a phone charger. Yet the brain sees, moves, speaks, remembers and learns.

Large AI systems, by contrast, are usually trained and run on many powerful computer chips in data centres, which need a lot of electricity and cooling. Brains and chips are very different things, so comparisons are rough. But the gap is large enough that engineers want to learn from brains.

Python: a rough average, just arithmeticpython

# Commonly cited estimates (rough): about 86 billion neurons, about 20 watts.
neurons = 86e9
watts = 20
per_neuron = watts / neurons
print(f"Average power per neuron: about {per_neuron:.1e} watts")
print(f"That is about {per_neuron * 1e9:.2f} nanowatts")

Idea 1: most neurons are quiet

In a brain, at any moment, only a small share of neurons are firing. Each spike costs energy, so fewer spikes means less energy. This is called sparse activity. In most AI models, by contrast, every neuron computes for every input.

Idea 2: work happens only when something changes

A normal computer chip ticks to a clock, doing work on every tick. Neurons are event-driven: a synapse does work mainly when a spike arrives. If nothing arrives, little happens.

Idea 3: memory sits next to computing

In most computers, memory and processor are separate. Data is carried back and forth all the time, and moving data costs a lot of energy, often more than the calculation itself. In a brain, the "memory" (synapse strengths) sits right where the "computing" happens.

Idea	Typical computer running AI	Brain
Activity	Dense: every unit computes every time	Sparse: few neurons fire at once
Timing	Clock-driven: work on every tick	Event-driven: work when spikes arrive
Memory	Separate; data moves back and forth	Stored in synapses, right where it is used
Precision	Exact digital numbers	Noisy, but good enough, and robust

A common myth

"The brain is just a slow computer"
"Brains and chips can be compared watt for watt exactly"

A fairer view

Brains and chips are built on very different principles, each good at different things
Comparisons are rough; they show a big gap and point to ideas worth borrowing

Real example

Think of a village street light that stays on all night, even when nobody walks by, and one with a motion sensor that switches on only when someone passes. Both light the path. The second uses much less power. Event-driven computing is the motion-sensor light.

Watch out

Be careful with exact claims such as "the brain is a million times more efficient than a GPU". Such numbers depend heavily on what you compare and how. Say "far more efficient for many tasks" unless you can cite a careful study.

Watch and read more

Lab

A toy count of operations for a dense network versus an event-driven spiking network, using made-up numbers, and the saving between them.

~20 min

The problem

This runs in your browser. All numbers are made up to show the idea; they are not measurements of any real brain or chip.

Run the starter. It makes a random spike record: spikes[t, i] is 1 if neuron i fired at step t. Each neuron sends FANOUT connections.
Set dense_ops: in a dense network, every connection does one operation at every step.
Set event_ops: in an event-driven network, only connections from neurons that fired do an operation.
Set saving = dense_ops / event_ops, and dense_energy and event_energy using the made-up ENERGY_PER_OP.
Change ACTIVITY to 0.2 and then 0.5. Write two lines on how the saving changes.

Practice lab

Edit the code and run it. Python runs in your browser — nothing is sent to our server.

Ctrl + Enter to run

Output

No output. Add print(...) to see values.

You are done when

dense_ops counts every connection at every step
event_ops counts only connections from neurons that spiked
saving and the two energies are computed from those counts
The note says the saving shrinks as activity rises, roughly 1 / ACTIVITY

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1At what activity level would the event-driven network do the same number of operations as the dense one? What does that tell you?Reveal

When every neuron fires at every step, that is ACTIVITY = 1.0, the counts are equal. In general the saving is about 1 / ACTIVITY. Event-driven computing pays off only when activity is sparse; if everything is active all the time, there is nothing to skip, and handling events may even add overhead.

Q2Our toy counts only operations. Name one real cost it leaves out.Reveal

It leaves out the energy of moving data, such as fetching weights from memory, which can be larger than the arithmetic itself. It also leaves out the cost of sending spike messages across a chip, keeping neurons' state up to date, and the power a chip uses even when idle. Real comparisons must measure these on real hardware.

Please sign in to continue.

Questions people ask

Is 20 watts a measured fact?

It is a commonly cited estimate for the human brain's power use, so we say "about 20 watts". The real value varies between people and moments.

If brains are so efficient, why not build AI like brains?

People are trying, with spiking networks and neuromorphic chips (lesson 7). But today's AI methods are tuned for normal chips such as GPUs, and training spiking networks is harder. Both paths are being explored.

Does sparse activity help normal AI too?

Yes. Some AI models use sparsity, for example by using only part of the network for each input, or by skipping zeros. The general idea "do not compute what you do not need" helps everywhere.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.