Nothing about a modern computer is conceptually new. The architecture was settled in the 1940s. What changed is the switch: the physical thing that turns a signal on and off. Every jump in computing power is really a jump in how small, fast and cheap that switch became. Tracking the switch through four generations tells the whole story.
| Generation | Switching technology | Consequence |
|---|---|---|
| 1st (1940s–50s) | Vacuum tubes | Room-sized, kilowatts of power, tubes burned out constantly |
| 2nd (1950s–60s) | Transistors | Smaller, cooler, far more reliable; no filament to fail |
| 3rd (1960s–70s) | Integrated circuits | Many transistors on one chip; cost per function collapses |
| 4th (1970s–now) | Microprocessors (VLSI) | A whole CPU on one chip; personal computers become possible |
Moore's law, stated properly
In 1965 Gordon Moore observed that the number of transistors on an economically produced chip was roughly doubling every year, later revised to about every two years. Notice what the claim is and is not. It is an observation about transistor count and manufacturing economics. It is not a law of physics, and it is not a promise that computers get twice as fast every two years — which matters, because since roughly 2005 the two have come apart.
That shift has a consequence students feel directly: extra cores only help if the software is written to use them. A program that does one long calculation in a single thread runs no faster on a 16-core chip than on a 4-core one. This is why parallel programming became a mainstream skill rather than a specialty.
The limits now being hit
- Physical size — current leading processes have features only a few dozen atoms across. Below that, electrons tunnel through barriers that are supposed to block them.
- Heat density — a modern CPU die dissipates more watts per square centimetre than a stovetop element.
- Cost — a leading-edge fabrication plant now costs well over ten billion dollars, so fewer companies can build one each generation.
- The memory wall — CPUs got faster far quicker than RAM did, so processors now spend much of their time waiting. This is why cache exists and keeps growing.
Where the industry went instead
- Parallelism — more cores, GPUs with thousands of small cores
- Specialized accelerators — neural processing units, video encoders
- Packaging — stacking chips vertically (3D), chiplets
- Performance per watt, which is what decides phone battery life
- Single-thread clock speed (roughly flat since 2005)
- Feature-size shrinks (slowing, each step costs far more)
- Price per transistor at the leading edge
- RAM latency, which has barely improved in two decades
One honest framing to carry into the rest of this course: the free performance improvements are largely over. For the last fifty years, a slow program could be fixed by waiting two years for faster hardware. That no longer works. Efficient code, appropriate hardware and good system design now matter more than they did for the entire previous history of the field.