From smartphones to AI, history has a habit of repeating itself, observes Bram Nauta.
How big would an iPhone 5 be if you built it with discrete transistors? I’ve asked my audience that question many times. Assume no wire leads, no printed circuit boards, no battery, no display. Just the black pea-sized plastic thingies. The technology of the 1970s.
People find it hard to imagine billions and billions of components. Think of a box shaped like an iPhone, filled with these black thingies – how big would that box be? A fridge? A car? A house? They have no clue.
And then I show a picture of the Eiffel Tower, with an inflated iPhone 5 standing next to it. The phone is almost as tall. And then I ask, what would the power supply look like? If every transistor needs a milliamp, and all those billions of transistors need power, the power requirements become huge: a few hundred megawatts. My next photo shows a small nuclear power plant next to the supersized iPhone 5. And note: an iPhone 5 is an old piece of junk that nobody wants to use anymore. The new ones contain even more transistors.
This example shows how insanely the technology has shrunk over the last decades. Imagine we’re back in 1970 and we’re making a plan to build this supercomputer with discrete transistors and put that nuclear power plant next to it. It would even be bigger, because we need more space than just those black peas. We need wires, power cables, racks with printed circuit boards, cooling and so on. That would be a ridiculous project, right? But thanks to technology scaling, we made it fit in our pockets.
Bizarrely, we’re now building such big iPhones. They aren’t standing up; they’re lying on their backs. They’re called data centers. And nuclear power plants next to them are no joke anymore; tech companies are buying retired reactors to power their data centers. We need these for AI.
AI benefits from scale; the more chips and racks, the better it becomes. Moore’s Law still helps, but performance increases come from more and more hardware in one place. The growth in computing is measured in hundreds of megawatts, just like my 1970 thought experiment.
AI answers many questions and can do amazing things people can’t. It has already affected how we grade our students, how we study and how we search for information. Suddenly, I can generate Python scripts that automate boring work. But the price we pay is sacrificing the landscape for concrete and copper constructions that drain our energy.
We, the chip people, have created little thingies that replace an Eiffel Tower-sized computer with something that fits in our pockets. But now, squeezing so many chips into data centers brings us back to that Eiffel Tower-sized computer.
While the iPhone 5 was a personal device, data centers are shared by millions of users. So, the energy needed per user remains modest. This brings us back to 1970, when computers filled large, air-conditioned rooms and were operated via dumb terminals. In the era of AI, our ‘smart’ phones are becoming the ‘dumb’ terminals.
