From sand to software: how a computer actually works

01 Silicon to wafer

Every computer begins as ordinary sand. The quartz in it is refined into metallurgical-grade silicon, then purified to 99.9999999% - a grade the industry calls "nine nines" - because even a handful of stray impurity atoms per hundred million is enough to spoil the tiny structures that make a chip work.

The purified silicon is melted, and a seed crystal is dipped into the melt and slowly withdrawn so the molten atoms lock onto it one layer at a time. This is the Czochralski process, and it grows a single flawless cylindrical ingot that can weigh more than a hundred kilograms. Diamond wire then slices the ingot into thin, mirror-smooth discs called wafers, the raw canvas of every chip.

Silicon wins the materials contest for two reasons. It is a semiconductor, meaning it can be switched between conducting and insulating behavior with small voltages, and it conveniently grows its own oxide: heat silicon in air and its surface turns into silicon dioxide, a stable insulator that fabrication lines have exploited for decades.

02 The transistor

A transistor is a semiconductor device that can amplify a signal or act as a very fast switch, and those two jobs underpin essentially all of modern electronics. As a switch it is the atom of computation; as an amplifier it is the workhorse of radios and sensors.

The workhorse of computing is the MOSFET, a voltage-controlled switch. Put a voltage on the gate terminal and a channel opens, letting current flow between source and drain; take the voltage away and the channel closes. It switches billions of times per second while its control terminal draws almost no steady current, which is what keeps chips from melting.

None of this is assembled by hand. Photolithography prints billions of transistors onto a single chip of silicon, layer by layer, the way a photograph projects an entire cityscape onto a postage stamp. A chip built this way - many components inseparably fabricated on one piece of semiconductor - is an integrated circuit.

03 Gates to processors

Wire a few transistors together in the right patterns and they begin to reason. Combinations of MOSFETs form logic gates - AND, OR, NOT - each one a tiny circuit whose output is a strict function of its inputs. From those primitives, all digital logic is constructed.

Gates cascade into adders that perform arithmetic on binary numbers, and into flip-flops, circuits that hold a single bit and can be clocked, which is the basis of registers and on-chip memory. A clock signal paces everything, so billions of switches change state in lockstep billions of times every second.

Assemble enough of this machinery to fetch instructions, decode them and carry them out, and you have a microprocessor: the entire central processing unit of a computer realized on a single integrated circuit. Transistor counts have grown so far, so fast, that the scale is hard to feel - the chart below plots it on a logarithmic axis, where every step is a tenfold jump.

04 The instruction set

Raw hardware understands only numbers, so something has to fix what the numbers mean. That agreement is the instruction set architecture, or ISA: the abstract interface that defines how software talks to hardware - which opcodes exist, how machine code is encoded, which registers are available and what each operation guarantees.

The ISA is a treaty with two signatories. Compiler writers ensure programs are emitted as legal instruction sequences; chip designers ensure their silicon executes every legal sequence correctly. As long as the treaty holds, software and hardware can each evolve independently - the program never needs to know what is physically inside the chip.

x86 and Arm are rival editions of that treaty. One descends from desktops and servers, the other from battery-powered devices, and their encodings differ enough that a binary built for one is gibberish to the other. That is why software is compiled per architecture, and why translating layers such as emulators have to intervene.

05 The storage hierarchy

Computers do not have one kind of memory; they have a pyramid of them. At the peak sit the CPU's registers, a few dozen slots of the fastest storage that exists. Below them come the L1 and L2 caches, then main memory (RAM), then SSD storage, then spinning disks. Each level down is vastly larger - and vastly slower.

The spread is enormous: a register answers in about 0.3 nanoseconds, L1 in roughly 1 ns, L2 in about 4 ns, RAM in around 100 ns, an SSD read near 0.1 milliseconds and a hard-disk seek near 10 milliseconds. The chart below puts those numbers on a log axis; on a human scale, if one CPU cycle took a second, RAM would feel like a five-minute wait and a disk seek like the better part of a year.

The fastest memory is also the smallest and sits closest to the execution units because physics charges a toll for capacity: signals travel finite distances, and dense storage is electrically slower to read and write. Caches and clever prediction exist to hide that gap - keeping the data a program is about to need as close to the processor as possible.

06 Operating systems and apps

For all its power, a processor still does one thing at a time, while a modern machine runs thousands of tasks at once. The operating system resolves that contradiction by multiplexing: a scheduler grants each process a slice of CPU time and switches between them thousands of times a second, fast enough that everything appears to run simultaneously.

The OS also gives every process its own private universe through virtual memory. Each program sees a clean address space of its own, which the hardware quietly maps onto physical RAM - and onto disk when RAM runs short. Isolation follows for free: one crashing program cannot scribble over another's memory.

Applications, meanwhile, are just portable sequences of instructions riding the ISA treaty. Compile a program once for a given instruction set and its logic runs on any machine that speaks the same dialect, with the operating system absorbing the differences in peripherals, memory and timing underneath. The same binary idea spans a phone, a laptop and a data-center server.

07 The limits

Half a century of progress is now pressing against physics. Packing more transistors into the same area concentrates heat faster than it can be removed; lithography is etching features measured in atoms, where a single misplaced boundary matters; and Dennard scaling - the old rule that let each generation shrink while staying cool and speeding up - has largely run its course.

Transistor counts keep climbing - NVIDIA's Rubin GPU carries about 336 billion of them in 2026 - but the payoff has changed shape. Extra transistors now buy parallelism, not per-core speed, and single-thread performance, the number users feel, improves by small percentages where it once doubled every couple of years.

So the frontier is rerouting rather than retreating: specialized accelerators for graphics, AI and video; 3D stacking that builds chips upward instead of outward; and chiplets, small dies connected inside one package, which trade a single vast monolith for a composed system. The tower of abstractions built from printed switches is not finished - it is being re-engineered from the ground up, again.