Why computers kept getting faster — and the wall

A phone in your pocket would outrun a room-sized machine from the 1970s — and out-think a supercomputer from the 1990s that filled a warehouse and drank a household’s worth of electricity. You did nothing to earn that. The machine just kept getting faster, year after year, for decades. Then, in recent years, that free ride quietly slowed. This lesson is about why computers raced ahead for so long, and the two hard walls — heat, and the size of an atom — that finally bit.

The doubling that ran for decades

Back at item 4 you met the transistor: a switch, built from a sliver of material that either lets electricity through or doesn’t. A chip is a city of them. The whole story of computer speed is one trick, repeated: engineers learned to make those switches smaller, every couple of years.

Smaller switches win three ways at once. You fit roughly twice as many on the same patch of silicon. Each smaller switch flips faster. And each uses less power. More switches, switching quicker, for less energy — that compounds into more work per second, on a chip no bigger than before, costing about the same.

This trend has a name: Moore’s law. It is worth being precise about what it actually says, because almost everyone gets it wrong. Moore’s law is an observation — really an economic prediction — that the number of transistors on a chip doubles about every two years. It is not a law of physics. Nothing in nature guarantees it. And it is about transistor count, not directly about speed. Speed rode along because smaller switches happened to be faster too, but the headline number is “twice as many switches.”

A worked example: feel the compounding

Doubling sounds gentle. It isn’t. Start with a chip holding a few thousand transistors — say 2,000. Now double it every two years and watch (the list jumps four years at a time, so each step is four times the one above):

• Year 0: 2,000
• Year 4: 8,000
• Year 8: 32,000
• Year 12: 128,000
• Year 16: about 500,000
• Year 20: about 2 million

Ten doublings — roughly twenty years — turn a few thousand into a couple of million. That’s a thousandfold leap (doubling ten times multiplies by about a thousand). Let it run another twenty years and you get another thousandfold — into the billions. A pocket chip today holds tens of billions of transistors, the harvest of about fifty years of this. Each one is a switch a few atoms across, flipping billions of times a second, and they nearly all behave. That is the most refined manufacturing humans have ever done, and for decades it got cheaper every cycle.

The first wall: heat

Every time a switch flips, it leaks a little heat. One flip, nothing. But take billions of switches, each flipping billions of times a second, packed into a chip the size of a thumbnail, and the heat per square centimetre climbs past a kitchen hotplate.

Here is the problem you can’t engineer around: you have to pull that heat out, or the chip cooks itself. And there is a limit to how fast you can move heat off a tiny surface — fans, metal, liquid, only do so much. So you cannot simply keep cranking the clock higher. (Clock speed, from item 7, is cycles per second — how often the chip steps forward.) Past a point, a faster clock makes more heat than you can shed. This is the main reason single-core clock speeds stopped climbing years ago and have largely sat still since. The engine hit its cooling limit.

The second wall: atoms

The other wall is even harder, because it’s set by nature, not by plumbing. The doubling depended on making switches smaller. But a transistor is now only a handful of atoms across. You cannot build a switch out of less than the atoms it’s made of — there’s a floor, and we’re near it.

Worse, down at that scale, electricity stops behaving politely. At a few atoms wide, charge starts to leak through barriers that should block it — a quantum effect, where the tiny world plays by spookier rules. The switch gets harder to keep firmly “off.” So shrinking still happens, but it’s slower, costlier, and fighting physics every step. The “smaller every two years” trick is running out of room.

The pivot: go wide, not fast

So the industry made a turn. If you can’t make one engine much faster, bolt on more engines. Instead of one ever-quicker processor (a core), chips now carry many cores — 4, 8, 16 — on one piece of silicon. Many cores can work in parallel: split a job into pieces and do them at the same time.

But here’s the honest catch, and it reaches back to item 10. Cores only help if the software is written to divide the work — and lots of jobs don’t split cleanly. Reading one sentence, then the next, has to happen in order; you can’t parallelise it. So “8 cores” is almost never “8× faster.” It’s “8× faster for work that happens to split,” and ordinary for everything else. That’s why a newer machine with more cores can feel barely quicker at the things you actually do.

The free lunch is over — and that’s the point

For decades, you got a faster computer by waiting. The chip doubled; your software sped up while you slept. That automatic gift is ending. Progress didn’t stop — it changed shape. From one engine that kept getting faster, to many engines you now have to learn to drive together. And underneath both sits a pair of floors that don’t move: the heat you can pull off a tiny surface, and the size of an atom.

It’s worth holding that lightly. The big number on the box — gigahertz, core count — never told the whole story, and tells less of it now. The next time a spec sheet dazzles you, remember the two walls underneath it. You’ll read one for real in the final lesson — cores and nanometres and all.