Power draw, to scale
Maximum ratings for the chips, the usual estimate for the brain.
The numbers
| System | Power | Against a 20 W brain |
|---|---|---|
| Human brain | about 20 W | 1 |
| NVIDIA H100 PCIe | up to 350 W | 17.5 times |
| NVIDIA H100 SXM | up to 700 W | 35 times |
| Frontier supercomputer | 21 MW | about 1 million times |
The chip figures are maximum ratings. The brain figure is a nominal budget derived from glucose uptake, usually given as 17 to 20 watts, and an adult body as a whole runs on about 100 watts.
A scale for AI training
The Neuroplatform paper says training GPT-3 required approximately 10 gigawatt-hours. At 20 watts, a brain would take about 57,000 years to use that much energy. The figure is an estimate quoted by that paper, so treat the result as an order of magnitude.
Why it is not a fair race
The two are doing different jobs. A GPU does exact arithmetic billions of times a second and a brain cannot multiply two six-digit numbers without effort. A brain runs a body, sees, walks and can recognise a face it has met once. Estimates of the brain at about 1 exaFLOPS are very rough, because neurons do not perform floating-point operations, and the authors who use that number say it serves only as an illustration.
Most of the brain's energy also does not go to what we would call computing. One PNAS paper finds communication uses about 35 times more energy than computation in the human cortex, and cortical grey matter is estimated at about 3 watts of the total.
Why researchers still care
Learning efficiency may matter more than power. The 2023 organoid intelligence paper notes humans learn a same-or-different task from about 10 samples, where machines failed at it with millions. That gap is the reason people are trying to build wetware at all.
Questions
How many watts does the human brain use?
About 20 watts is the usual figure, with estimates of 17 to 20 watts from glucose uptake.
How much power does an NVIDIA H100 use?
The SXM version is rated at up to 700 watts and the PCIe version at 300 to 350 watts.
Is the brain more efficient than a GPU?
On power, yes by a wide margin, but they do different work. The comparison is an illustration, not a benchmark.
Sources
- Balasubramanian, Brain power, PNAS, 2021
- Levy and Calvert, Communication consumes 35 times more energy than computation in the human cortex, PNAS, 2021
- NVIDIA H100 specifications, University of Tübingen ML Cloud
- Smirnova et al., Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish, Frontiers in Science, 2023
- Jordan et al., Open and remotely accessible Neuroplatform for research in wetware computing, Frontiers in Artificial Intelligence, 2024