A high-end CPU these days uses nearly 100W of power. But it doesnâ€™t have nearly the computing power of a human brain. AI is in good part, perhaps mainly, a software problem, but raw computing power seems lacking too. So how many of todayâ€™s CPUs would it take to build a computer with human intelligence? Say at least 10,000 Opterons.
This 10K CPU computer system would use 1MW of power. So how does that compare to a human brain? A person runs on 2000 Kcal / day.
At 86400 s / day this is 23 cal / s x 4.187 calories / J = 97 J / s. J / s are equal to Watts so this is 97 W.
Say 40% of the bodyâ€™s energy is used by the brain. Then a personâ€™s brain uses 40 W, as much a weak light bulb. Which is order-of-magnitude correctâ€“your head is a little cooler than a weak bulb but the bulb is smaller.
So a human intelligent computer would use 1MW of power while a personâ€™s brain uses 40W. The human brain is 25,000X more effcient than a computer. This says a few things about todayâ€™s computers. They are terribly inefficient, at least for the types of computations an AI needs to do. And todayâ€™s AI software design doesnâ€™t capture the organization of biological computers. We have 1000 CPU systems today, and could build 10K CPU systems. But no system today is as clever as a mouse. Todayâ€™s AI may have crested the house fly brain goal post. But the lack is clearest in the hardware architecture.