How many transistors are in a processor: from 2,300 to 20 billion
"How many transistors are in a processor" is the question beginners ask most often, and it has an exact answer: the Intel 4004, the first processor sold commercially in 1971, had 2,300 transistors, while modern chips are counted in tens of billions. In fifty years that is a growth of about ten million times. Below is how the numbers changed, what a die is actually made of, why there are more transistors than there is work to do, and how many the processor you build in this course would need. The road from a transistor to a processor runs through four skills, collected on the Circuit Design page.
How the numbers changed
| Processor | Year | Transistors | What it was |
|---|---|---|---|
| Intel 4004 | 1971 | 2,300 | 4-bit, 16 instructions, built to order |
| Intel 8086 | 1978 | 29,000 | 16-bit, the basis of the IBM PC |
| Intel 80386 | 1985 | 275,000 | 32-bit with virtual memory |
| Intel Pentium | 1993 | 3,100,000 | Pipelined, two execution units |
| Intel Pentium 4 | 2000 | 42,000,000 | Frequency mattered more than power |
| Intel Core 2 | 2006 | 291,000,000 | Two cores, 65 nm process |
| Intel Haswell | 2013 | 1,420,000,000 | 14 nm, more cache, integrated graphics |
| Apple M1 | 2020 | 16,000,000,000 | A 119 mm² die, 5 nm |
| Apple M2 | 2023 | 20,000,000,000 | The next generation, the same approach |
Note what the axis means: each step along the chart is a tenfold increase, so bars at the same distance differ in volume by a factor of ten.
Moore's law and where it stopped
In 1965 Gordon Moore noticed that the number of transistors on a die grows predictably, and he stated the observation that became known as Moore's law: roughly every two years chip capacity doubles. For fifty years the curve followed that rule almost literally, and the whole industry was built on it.
After that the rule started working halfway. Transistors kept being added, but doubling stopped meaning doubling the speed: frequency is limited by heat dissipation, and transistors had to become three-dimensional to fit more of them onto the same area. Since 2015 what gets doubled more often is not the whole die but its parts: several chips on one substrate and vertical stacks of memory. The twenty billion transistors in the Apple M2 are two dies side by side, not one.
One calculation shows where that figure comes from. The M1 die is 119 mm² and holds 16 billion transistors, a density of about 134 million transistors per square millimetre. Fitting 20 billion transistors at that density takes roughly 150 mm²: the area of a small fingernail, and it still does not fit in the head without a calculator.
What a die is made of
The smallest share belongs to what people usually want to see: arithmetic and logic. Most of the volume goes to memory.
| Part of the die | How it is built | Share of volume |
|---|---|---|
| Cache memory | Static memory: one bit takes several transistors | Most transistors in a core |
| Surrounding main memory | Separate DRAM dies, sometimes stacked vertically | Counted apart from the core |
| Logic and arithmetic | Gates built from field-effect transistors | Tens of percent |
| Analog part | Generators, comparators, converters | Small in volume, mandatory in practice |
| Input and output | Signal drivers, level converters | Small, but they decide whether the board works at all |
The reason is that a memory element costs more transistors than a logic gate does. Holding a single bit of static memory takes no fewer than six transistors, and each of them burns current continuously to keep its state. Where fast access to data matters, it pays to spend transistors on such memory rather than on extra logic.
More transistors does not mean more work
The common explanation says a processor uses one percent of its transistors. Research by Karl Fritz of the University of Erlangen-Nuremberg, published in IEEE Micro in 2017, showed a different picture: in a Haswell processor about a fifth of the transistors are idle at any given moment, because the area is there but the running program does not reach them.
The cause lies in the design principle itself: a complex processor is designed for the worst case, and a program almost never loads every block at once. That is where the sleeping transistors come from, along with the power-saving states and the modes where a whole chip runs at a reduced clock. A running processor uses most of its area, but not all of it at the same time.
What the course processor would need
The teaching processor with 15 instructions, 8-bit registers, and Harvard memory belongs to the class that was built on NMOS transistors in the late 1970s. For comparison: the 6502 had 3,510 transistors, the Z80 had 8,486. Both chips could do what the processor in this course does: arithmetic, comparison, branching, and memory addressed by location.
The order of magnitude follows from the design. A register is a few transistors, an adder is tens, a multiplexer and a decoder are hundreds, and a memory block is thousands per bit. The course processor amounts to a few thousand transistors, three orders of magnitude below a modern core. The gap in scale is explained not by the difficulty of the ideas but by what the factory could do: in 1971 the same idea simply did not fit on a die.
How many transistors did the first processor have?
Intel 4004, released in 1971, had 2,300 transistors. It was a 4-bit processor with 16 instructions aimed at calculators, and it was built to order for a single Japanese company, Busicom.
Why does a processor need so many transistors if it performs simple operations?
Most of them go into cache memory: a bit of static memory takes several transistors, and current processors hold tens of megabytes of cache. The second largest share is several cores connected to shared cache, a bus, and power controllers.
Do all transistors in a processor run?
No. Research by Karl Fritz, published in IEEE Micro in 2017, showed that about 20 % of the transistors in a Haswell processor are idle at any given moment: the area is there, but the running program does not reach them.
How many transistors would the processor from this course need?
Processors of this class were built on NMOS in the late 1970s: the 6502 had 3,510 transistors, the Z80 had 8,486. The course teaching processor with 15 instructions and 8-bit registers belongs to the same class, so the figure is thousands of transistors rather than billions.
Check yourself
Why does one bit of memory cost more than one gate?
A gate computes a value from its inputs and does not have to remember the result. A static memory element holds its state between clock cycles, so it must keep it while power is applied: two transistors for storage plus the access transistors. Six transistors per bit instead of one is the normal price of on-die memory.
What came before the transistor?
Before the transistor, logic was built from electromechanical relays, and later from vacuum tubes. A relay computer from the 1940s filled a room, consumed kilowatts, and took seconds per operation. The transistor, in 1947, replaced a unit the size of a crate with one the size of a thumbnail, and the number of such units per die started growing along what became Moore's law.
Why can't we call a modern processor "millions of times faster" than the 4004?
There are ten million times more transistors, but the clock rate grew only a few thousand times; the rest came from pipelines, caches, and parallelism. Processor performance is measured in operations per second, not in transistors, and per unit area transistors now work thousands of times more efficiently.
How a transistor turns into a logic gate is covered in from transistor to gate, how the whole processor is put together in how a processor works, and how a whole board rather than just a core is assembled in the Tang Nano 9K, from ZIP to board.