Co-packaged optics
Moving data by light instead of copper, with photonic engines placed beside the processor to cut the power of chip-to-chip links in AI systems.
A modern processor is a sliver of purified sand carrying billions of switches, wired together by a dozen or more layers of copper and patterned with light just 13.5 nanometers long. This guide lets you zoom into one, run one, heat one up, and build one, a step at a time.
The shimmer is real physics: each color is computed from thin-film interference in the oxide layers on silicon. Move the light to tilt the wafer.
A microprocessor is a complete central processing unit on a single chip. It repeats one loop billions of times a second: fetch an instruction from memory, decode it, execute it, and store the result. Everything else on this page is about making that loop faster, smaller, cheaper and cooler.
The processor familySame transistors, different organization for different jobs
Fall from a CPU package into its silicon atoms. Along the way, compare what you see with things of the same size: a hair, a blood cell, a virus, a strand of DNA.
Powers-of-ten zoomEvery level is drawn to scale; the lane on the right shows a familiar object at the same scale
to count the transistors in NVIDIA’s Rubin GPU at one per second (336 billion).
If a 48 nm gate pitch were scaled up to one meter, a human hair (~75 µm) would be about 1.6 km wide.
transistors had been made by 2018, the most manufactured human artifact in history.
Source: Computer History Museum
1.15 trillion semiconductor units shipped in 2021 alone.
Source: ASML
Until the mid-1990s a process name like “350 nm” matched the gate length. Then the numbers drifted into marketing labels: no feature on a “3 nm” chip measures 3 nm. Real density comes from the contacted gate pitch and the metal pitch. Pick a process to see them drawn to the same scale as its name.
Sources: 2 nm process and 3 nm process comparison tables (foundry disclosures and IRDS projections); IEEE Spectrum for Intel 22 nm.
In 1965 Gordon Moore predicted that the number of components per chip would keep doubling; in 1975 he settled on every two years. Adjust the doubling period and see how well it fits real chips.
Sources: Transistor count (Wikipedia), SiliconANGLE (Rubin), Crypto Briefing (WSE-3).
Every processor is built from MOSFETs. A voltage on the gate pulls electrons into a thin layer of silicon underneath it. Above the threshold voltage that layer connects source to drain and current flows: the switch is on. Below it, only a small leakage current trickles through.
Transistor simulatorAn n-type transistor; current per micrometer of width
The better the gate surrounds the channel, the more completely it can switch it off. Planar transistors control the channel from one side; FinFETs from three; gate-all-around nanosheets from all four. That shows up as a steeper curve below threshold and far less leakage, which matters when a chip has billions of transistors idling at once.
At room temperature, conventional transistors cannot turn off faster than about 60 millivolts of gate voltage per tenfold drop in current. This “Boltzmann limit” follows from the thermal energy of electrons. It is one reason supply voltages stopped falling, and why engineers fight for every millivolt.
Pair n-type and p-type transistors and you get CMOS: one network pulls the output up to the supply, the other pulls it down to ground, and almost no current flows while nothing changes. Four transistors make a NAND gate, and NAND gates alone can build any digital circuit.
CMOS gate, transistor by transistorClick the inputs; lit transistors are conducting
Amber path: output tied to the supply (1). Cyan path: output tied to ground (0). In steady state no current flows through either; current only moves while the output switches.
A 4-bit adderFlip the bits and watch the carry ripple from right to left
Sum: –Delay so far: . Each full adder needs about two gate delays to pass its carry on, so a 64-bit ripple adder would be far too slow. Real CPUs use carry-lookahead adders that compute all carries in parallel. Click a full adder to see inside it:
This is a working 8-bit processor with an accumulator, a program counter, 64 bytes of memory and 14 instructions. Step one clock tick at a time to watch each instruction get fetched, decoded and executed across the internal bus. Then write your own program.
Teaching CPU sandboxEach tick moves one value over the bus
Cyan outline: program counter. Amber: memory address register. Brighter text: program code.
Every instruction starts the same way: the program counter’s address goes to memory, the instruction comes back into the instruction register, and the counter moves on. The control unit decodes the opcode and then drives a short sequence of transfers specific to that instruction.
A modern core runs the same loop, but overlaps dozens of instructions, executes several per clock, predicts branches, and keeps hundreds of instructions in flight. The next two sections show what that looks like in silicon and in time.
A representative hybrid processor: performance cores for fast single threads, efficiency cores for background work, a shared cache, a GPU, an NPU, and the interfaces that feed them. Choose a workload and a cooler. The simulation sets clock speeds with DVFS, computes each block’s power, and solves the heat flow through the silicon in real time.
Cores sit near the cache they share, the memory controllers sit at the edge where the memory bus leaves the package, and blocks that are rarely busy together can share the power budget. Idle blocks are power-gated: switched off completely so they stop leaking.
A performance core at full boost can dissipate several watts per square millimeter, more than a stove hotplate. The cooler sets how long a chip can sustain its top clock; power limits and thermal throttling decide the rest. Notice how the hottest point sits on the busiest execution units.
Clock speeds stopped climbing in the mid-2000s. Since then, speed has come from doing more per tick: pipelining, issuing several instructions at once, predicting branches, keeping data in caches close to the core, and adding cores.
An instruction set architecture defines the instructions a processor understands. Here is the same line of C, a = b + c;, in the three that matter today. x86 is a CISC design whose arithmetic can read memory directly; Arm and RISC-V are RISC designs where only loads and stores touch memory.
Same task, three instruction setsReal machine-code bytes on the right
A five-stage pipeline overlaps instructions like an assembly line: fetch (IF), decode (ID), execute (EX), memory (MEM), write-back (WB). Dependencies between instructions force stalls; forwarding results straight from one stage to another removes most of them. Branches are resolved late, so a wrong guess flushes the instructions fetched behind it.
Pipeline simulatorRows are instructions, columns are clock cycles
Main memory is hundreds of clock cycles away, so cores keep recently used data in small caches. Each address is split into a tag, a set index and a byte offset. A cache with more “ways” per set suffers fewer conflicts. This one has eight 16-byte lines.
Cache simulatorLRU replacement; hit = 4 cycles, miss = 100 cycles
Try this: “Three addresses that collide” is all misses when direct-mapped, still all misses at 2-way (three lines fight over two slots), and nearly all hits at 4-way.
Adding cores only speeds up the part of a program that can run in parallel. In 1967 Gene Amdahl pointed out that the serial remainder caps the gain: speed-up = 1 / ((1 − p) + p / N). That is why every modern chip still needs at least a few very fast cores.
Before any silicon is touched, a processor exists as code. Engineers describe it in a hardware description language, simulate it for billions of cycles, and let EDA software turn it into the exact shapes on every layer. A new high-end processor takes large teams several years. Follow one full adder through the flow.
A leading-edge wafer spends about three months in the fab. One GlobalFoundries fab described roughly 700 steps building a 60-to-75-layer “cake”. The same short loop repeats layer after layer: deposit, coat, expose, develop, etch or implant, clean. Walk the line station by station.
Fabs filter air to ISO class 5 or better; ISO 1 allows at most 10 particles of 0.1 µm per cubic meter. City air holds about 35 million particles of 0.5 µm per cubic meter.
Source: RS Components, Cleanroom
and about three months for one wafer at a GlobalFoundries fab; the most advanced logic needs more.
Source: AP via The Spokesman-Review
of water a day for a large fab, much of it ultrapure; fabs increasingly recycle it.
Source: Area Development
reported price of one 2 nm-class wafer from TSMC, about double a 5 nm wafer.
Source: Tom’s Hardware
Chips are built by repeating a handful of operations in the right order. Pick a goal and try to reach it. Masks let light through their bright parts; this resist is positive, so exposed areas dissolve when developed.
Process builderA cross-section of the wafer surface, 48 columns wide
Each generation needs breakthroughs on several fronts at once: shorter-wavelength light, new transistor shapes, new materials, and new ways to stack and connect dies.
The smallest half-pitch a scanner can print is roughly k₁ × λ / NA: shorter wavelength λ, larger numerical aperture NA, and cleverer process tricks (smaller k₁, physically no lower than 0.25) all help. Below the limit, the image blurs into gray. Multiple patterning splits a dense pattern over several exposures, at the cost of more steps.
Wavelengths: ASML. EUV resolves about 13 nm at 0.33 NA and about 8 nm at 0.55 NA: Tom’s Hardware.
As single dies approach the size limit of a lithography exposure (about 26 × 33 mm), performance increasingly comes from packaging several dies together. Chiplets, silicon interposers and stacked HBM memory are why AI accelerators can reach hundreds of billions of transistors.
Traditionally, power and signals share the dozen-plus metal layers above the transistors, so supply current must thread down through all of them. Backside power moves the power network under the thinned silicon, feeding transistors from below. Intel shipped it first with PowerVia on 18A; TSMC’s A16 “Super Power Rail” is planned for the second half of 2026.
The result is less voltage droop, more room for signal wires, and denser cells. See it in the anatomy simulator: choose “Cross-section” and switch between front-side and backside power.
Early chips needed a handful of elements. A modern process uses highlighted here, from hafnium gate insulators to the tin that makes EUV light. Hover or tap an element.
Switching power follows P ≈ C × V² × f. Until the mid-2000s each shrink also lowered the voltage (Dennard scaling), so clocks rose for free. When leakage stopped voltages from falling, higher clocks meant steeply higher power, and the industry turned to more cores and specialized accelerators.
Voltage, frequency and powerHigher clocks need higher voltage, so power rises much faster than speed
That last line is the whole case for multicore chips: two slower cores can do the same work for a fraction of the energy, if the software can split it.
Data: Karl Rupp, microprocessor trend data (CC BY 4.0), from M. Horowitz, F. Labonte, O. Shacham, K. Olukotun, L. Hammond and C. Batten; covers chips through 2021.
The world bought about $796 billion of semiconductors in 2025. In June 2026, WSTS forecast $1.51 trillion for 2026, almost double, driven by AI data centers and a roughly 250% jump in memory sales. Its August update, using second-quarter actuals, pointed to about $1.7 trillion.
US$ per year; striped bars are WSTS forecasts
Sources: WSTS and SIA annual releases; WSTS Spring 2026 forecast (2025 actual, 2026–27 forecast); WSTS Q2 2026 update.
WSTS product groups and regions. Microprocessors and microcontrollers fall under “Micro”; much of the GPU and AI-accelerator boom shows up under Logic.
Source: WSTS Spring 2026 forecast. Percentages are change from the previous year.
Source: TrendForce, share of top-10 foundry revenue (US$53.5 billion in the quarter).
In 2022, nearly all leading-edge logic capacity (below 10 nm) was in Taiwan and South Korea. Governments responded: the US CHIPS and Science Act committed $52.7 billion, and the EU Chips Act aims to mobilize more than €43 billion. The SIA and BCG projected in 2024 that the US share of leading-edge logic capacity would rise from almost zero to 28% by 2032. Export controls on advanced chips and tools have also reshaped where chips can be sold and built.
Source: SIA/BCG, May 2024
Follow one chip around the worldNo country makes a leading-edge processor alone
Under the India Semiconductor Mission, Micron opened a $2.75 billion assembly and test plant in Sanand, Gujarat, in February 2026, and Kaynes Semicon followed in March. Tata Electronics is building the country’s first commercial wafer fab at Dholera with Taiwan’s PSMC, aimed at mature 28–110 nm processes and about 50,000 wafer starts a month, with full production targeted around 2028.
Sources: Site Selection, Business Today
Fabs are among the most resource-intensive factories on Earth. Producing a large chip on a 300 mm wafer can take about 8,300 liters of water, including some 5,700 liters of ultrapure water, and EUV tools and cleanroom air handling draw large amounts of electricity. Fabs also use fluorinated gases with high warming potential, so abatement and water recycling are now major engineering priorities.
Source: Veolia
Few companies do everything. Integrated device manufacturers design and build their own chips; fabless companies design and outsource manufacturing to a foundry; IP companies license instruction sets and cores; equipment and software makers supply everyone.
Source: Gartner, January 2026 (preliminary). Total market $793 billion.
Arm-based chips are rising alongside x86: Arm says its CPUs now handle about half of the compute at the largest cloud providers.
Sources: Mercury Research via HotHardware, Guru3D, TweakTown; Arm.
The roadmap runs into angstroms (tenths of a nanometer), but the labels are still generation names. Progress now comes from stacking: transistors on top of transistors, power wiring underneath, and dies bonded on top of dies. In June 2026 IBM showed a 0.7 nm-class “nanostack” research chip with nearly 100 billion transistors on a fingernail-sized die.
Moving data by light instead of copper, with photonic engines placed beside the processor to cut the power of chip-to-chip links in AI systems.
Glass package cores are flatter and more stable than organic ones, allowing larger multi-die packages with finer wiring.
Copper pads bonded directly at micron-scale pitch let memory and logic be stacked with thousands of times more connections than solder bumps.
Quantum processors, neuromorphic chips and in-memory computing target specific problems; for general-purpose computing, CMOS is expected to dominate for decades.
Sources: imec roadmap via Tom’s Hardware, imec, MIT Technology Review (IBM), AnandTech (A16), Tom’s Hardware (High-NA plans).
Two sandboxes that tie the guide together: the economics of yield, and the trade-offs of designing your own processor. Both use simplified models with their assumptions stated.
Yield and cost labHow die size and defects set the price of a chip
Chip architectSpend a silicon budget and see what you get
Twelve questions drawn from the sections above. Each answer comes with an explanation.
Figures reflect sources available in November 2026. Market forecasts are revised often; check the linked releases for the latest numbers.