Human Brain vs Artificial Intelligence: Which Uses More Power?

Human brain vs artificial intelligence, by the numbers: the brain runs on about 20 watts, while one NVIDIA H100 chip is rated at up to 700 watts, about 35 times as much by our arithmetic. The brain has about 86.1 billion neurons. Meta's largest open Llama 3 model has 405 billion parameters and was trained on up to 16,000 of those chips. These numbers measure hardware, not intelligence.

Below: the power gap with every figure tied to a paper or spec sheet, what a neuron count can and cannot be compared with, how fast each side is, and why none of it settles which is smarter. Our own arithmetic is marked.

Is the human brain like a computer?

In some ways, yes, and the comparison has limits. Both turn energy into signals, and both can be measured in watts. But the human brain as a computer is a metaphor, not a blueprint. A synapse is not a storage cell and a memory is not a file, as our guide to how much data the brain can store explains. A model's parameter is not a synapse either, a point our research study on AI size compared with a brain covers in detail.

People who search for the human brain and computers often expect a computer vs human brain scoreboard. The useful question is narrower: what can be measured on both sides, and what does each number say? The rest of this page takes the measurable ones: power, counts and speed.

Human brain vs artificial intelligence: how much power does each use?

The brain uses about 20 watts, and a single AI chip is rated at up to 700. Balasubramanian (PNAS, 2021) gives the brain's energy budget as about 20 watts, with 17 to 20 watts derived from glucose uptake. The brain is about 2% of body weight but uses about 20% of the body's energy (Raichle and Gusnard, PNAS, 2002). NVIDIA's H100 page lists a thermal design power (TDP) of up to 700 watts for the SXM card and 350 to 400 for the NVL card.

• Human brain: about 20 W, or 0.48 kilowatt-hours over a day (our arithmetic)
• NVIDIA H100 SXM: up to 700 W, 35 times the brain at 20 W and 41 times at 17 W (our arithmetic)
• NVIDIA H100 NVL: 350 to 400 W, 17.5 to 20 times the brain at 20 W (our arithmetic)
• Llama 3 405B training: up to 16,000 H100 chips at 700 W each, 11.2 megawatts at the chips' rating, about 560,000 times a 20 W brain (our arithmetic)

The last row comes from Meta's Llama 3 paper, which says the largest model was trained on "up to 16K H100 GPUs, each running at 700W TDP". We multiplied the two figures. Read it as a ceiling for the chips alone. A TDP is a rating, not a measured draw, and the sum leaves out servers, networking and cooling.

Over a full day the gap is easier to picture. A brain at 20 watts uses 0.48 kilowatt-hours. One H100 SXM card at its 700-watt rating uses 16.8 kilowatt-hours (our arithmetic, assuming it ran at its rating all day, which real chips do not). Both figures are power, not thinking: no one has put a kilowatt-hour value on a thought.

How many neurons does the brain have, and what is the equivalent in an AI model?

The human brain has about 86.1 billion neurons, and an AI model has no neurons at all. Azevedo and colleagues (Journal of Comparative Neurology, 2009) counted 86.1 billion neurons, plus or minus 8.1, in the average adult male brain, with 84.6 billion non-neuronal cells. Only 19% of the neurons sit in the cerebral cortex, about 16 billion by our arithmetic, as our guide to human brain vs dog brain shows.

What a model has is parameters, the numbers it adjusts during training. Meta's largest Llama 3 model has 405 billion. That is 4.7 parameters for each human neuron (our arithmetic), and the number means nothing on its own. A neuron is a living cell that fires signals, and a parameter is a stored number. Counting one against the other is like comparing a village's population with its library's page count.

Energy shapes the brain's design. Lennie (Current Biology, 2003) estimated that the high cost of nerve signals limits the share of cortical neurons that can be strongly active at once to possibly fewer than 1%. Our guide to how much of the brain we use reads that estimate and the 10% myth around it. See also our human brain facts.

Is a computer faster than the brain, and at what?

At raw arithmetic, a chip is faster by many orders of magnitude. NVIDIA's table lists 34 teraFLOPS of double-precision (FP64) arithmetic for the H100 SXM, which is 34 trillion operations a second. A calibrated sample of 1,469 adults averaged 231 milliseconds (raw) to react to a signal, as our reaction time guide reports. In that time one H100 completes about 7.85 trillion operations (our arithmetic).

Memory speed tells a similar story. NVIDIA lists 80 GB of memory and 3.35 terabytes a second of memory bandwidth for the H100 SXM. The brain's memory is not a separate bank with a bandwidth to quote: a memory is spread over many synapses and rebuilt each time you recall it, as our storage guide notes.

That comparison is a trick of units. A reaction is a sensing, deciding and moving task, not arithmetic, and no common unit measures understanding. The computer and human brain also differ in what each does with its time: the chip performs one narrow kind of operation very fast, while the brain handles perception, movement, memory and language on 20 watts. Speed in operations per second says little about either.

Can AI match the brain's efficiency, and is a computer smarter than a human brain?

Not on the evidence we read. Efficiency means work done per watt, and no source we read gives the energy cost of the same task for a brain and a chip. The 35-fold gap compares what each device draws, not what it accomplishes. A chip rated at 700 watts could be less or more efficient than a brain on a given job, and nothing here says which.

Is a computer smarter than a human brain? Power, neuron counts and speed describe the hardware. Intelligence is measured on tasks, and on those the answer changes with the task. Our guide is AI smarter than humans compares ability, what is AGI explains the target many labs name, and types of artificial intelligence shows how researchers grade it. People who ask about AI and human brain comparisons often want one winner. The data give three different scoreboards.

IQ tests exist to place one person against others, not to rank machines. If you want to see how a human score maps to a percentile, our IQ percentile calculator does the conversion. For a famous mind in the history of computing, see the Alan Turing profile.

To see how you score against other people, take the IQ Metrics test.

Common questions