Products You May Like
A CPU or central processing unit is a computer system’s most powerful and important component. CPUs are integrated circuits: solid-state components found in devices of all kinds, from phones and laptops to PCs and even servers in data centers.
What Does a CPU Do?
CPUs are often compared with brains, as it is the CPU’s job to wrangle different streams of load-bearing data moving through the system.
In a computer, the CPU is the generalist, built to handle the broadest variety of data. At the risk of being a bit reductive, CPUs are good at doing many different things simultaneously (reading, writing, fetching, math, and logic), while GPUs are better at doing the same operation (e.g., a transform) to many different data points.
CPUs perform a complex juggling act, including tasks such as branch prediction and speculative execution that must be coordinated in time. This makes a CPU capable of handling nearly any code compatible with its instruction set architecture.
Anatomy of a CPU
Physically, a CPU is a slice of crystalline silicon. It’s laser-etched with special functional designs and then sealed inside a metal housing, which protects the sliver of crystal and helps it shed excess heat.
Credit: Olivier Collet/Unsplash
In a PC, the CPU is socketed into a motherboard, while in laptops or mobile devices, it may be soldered into place.
Credit: Francesco Vantini/Unsplash
Every semiconductor company has its own flagship designs, but the industry has undergone some convergent evolution. Inside any given CPU, you’ll find a version of the same few common parts.
1. Control unit
Like an overwatch, the control unit’s job is to direct the flow of information within and between the CPU and other elements. It decodes information received from memory, converting the data stream into operations that fit its instruction set.
2. Registers
With all the data moving through a CPU at any given time, the processor needs to have a way to stage data that’s fast and easy to access. To keep critical information close at hand, CPUs use a kind of memory called registers.
Registers are physical memory defined by their size (given in widths such as 8 bits, 16 bits, 32 bits, and so on). For specialized instructions like AVX-512, CPUs use 512-bit-wide registers. Sometimes, the units are given as bus widths, which is a reference to bussers: restaurant staff whose job it is to quickly clear or “bus” tables and get the dishes to the kitchen sink. These unsung heroes carry tableware using a tub or tray.
CPUs keep data in various types of memory, such as buffers, caches, and registers. (Fun fact: In many block diagrams, you’ll see “cache” abbreviated “$” for brevity.) Memory management units help to fetch, coordinate, and deliver data that the CPU needs to keep in play.
3. Core(s)
A CPU core is like a reactor core: It’s where the work gets done. Most CPUs these days have between two and eight cores, which operate at frequencies given in hertz (Hz—one Hz is one cycle per second, a megahertz is a million cycles per second, a gigahertz is a billion, and so on). Each core is responsible for one or more data streams called threads.
CPU cores comprise multiple subunits, including ALUs (arithmetic logic units) and FPUs (floating point units). In this block diagram of the AMD Zen 5 desktop version below, you can trace the flow of information from the “frontend,” which includes the control unit, through ALUs into the L1 cache.
Credit: Chips and Cheese
An ALU’s job is to handle arithmetic and simple Boolean logic (i.e. AND, NOT, and OR operations). The ALU receives information from the registers, processes it according to instructions from the controller, and then stages it in an output buffer. ALUs are the basic building blocks of many types of processors, including CPUs and GPUs.
In comparison with the ALU, floating point units or FPUs handle more complex operations, chiefly decimal math. Floating point operations are so important for precision work that they lent their name to a unit of computing throughput: the root of words like gigaflop and teraflop is the acronym FLOPS, for FLoating-point Operations Per Second.
4. Clock
If the control element directs the course of information, the clock dictates its tempo. Not unlike the timing chain in a car, the entire CPU coordinates itself based on the clock. The timing chain keeps everything physically in sync no matter how fast the engine is running. Likewise, a CPU’s clock speed can vary, but whatever the clock is set to, the system runs at that frequency.
How CPUs Are Made
Once upon a time, the processing units that powered room-sized computers like ENIAC and UNIVAC were built from macroscale materials like wires and resistors, by human technicians. As transistors shrank, it became flat-out impossible for a human to do the work of finescale production by hand.
Modern chips are made using a beam of high-energy UV light, in a process called photolithography.
To make a CPU, one begins with a great big hunk of monocrystalline silicon, known as an ingot or boule. Boules are grown in cleanrooms by vapor deposition and laser-cut into discs called wafers. Then, each wafer gets a kind of stencil called a mask, a physical pattern that allows light to reach only certain areas.
For some CPUs, features on the silicon are so tiny that the wafer has to be immersed in a layer of liquid to focus the light properly. But because even single-atom flaws in the silicon can disrupt the nanoscale features on a chip, manufacturers often choose to grow the boule on a substrate: a slab of flawless created sapphire.
Credit: TSMC
After etching, most wafers go under a different laser that cuts the wafer into many individual units, which are then encased in a metal housing. The silicon, housing, and any pins or electrical contacts collectively make up a CPU’s package, often simply called a chip.
Why Silicon?
Silicon is the material of choice for most CPUs because of its desirable electrical properties. Silicon is a semiconductor: neither an insulator (like a noble gas) nor a fully fledged conductor like copper wire. Electrons move through silicon rapidly, but in ways we can control.
Like carbon, silicon has four valence electrons and predictable behavior when exposed to electrical current. Adding trivalent or pentavalent atoms to the silicon vapor changes the electrical properties of the resulting crystal—and when two such regions are laid down side by side in stripes, they form the material basis of a metal-oxide semiconductor field effect transistor, or MOSFET.
Notable CPU Makers
Broadly speaking, there are two main classes of consumer CPU, divided by their instruction set architecture (ISA): x86, and ARM. Intel’s Core i5, i7, i9 and their kin are x86 CPUs, as are AMD’s Ryzen CPUs. ARM CPUs are used in smartphones with some limited presence in data centers and supercomputing.
Apple notably switched to ARM processors for its desktops and laptops several years ago. Qualcomm aims to follow suit with its Snapdragon X Elite processors, used in some Windows PCs. There’s also an open-source ISA known as RISC-V (pronounced risk-five). RISC-V is an emerging platform of choice for many embedded projects, particularly those wishing to take advantage of an open-source CPU ISA.
Some semiconductor companies design CPUs and build them in-house, at their own foundries. During the aviation age, in the heyday of early computing, there were a surprising number of different foundries operated by the likes of IBM, Fujitsu, and Toshiba, in addition to familiar names like Intel and Samsung. Today, it’s basically Samsung and Intel.
CPU foundries are outlandishly, hilariously expensive to build and operate. So to turn a profit, a company has to have deep pockets and guaranteed demand. A foundry can draw as much power as a small city. The industrial machines required to build CPUs are delicate instruments built by other, still more exacting machines (the latter produced by basically two companies, Applied Materials and ASML). Consequently there are only a few dozen of these facilities on the face of the Earth.
Semiconductor companies like AMD, Apple, Qualcomm, and Nvidia are known as “fabless” firms because they design CPUs but don’t build them. Fabless firms may work with so-called “pure-play” foundries, which only make chips for other customers. Pure-play foundries don’t build any of their own designs, and they aren’t in direct market competition with their customers.
Many, if not most, fabless semiconductor companies have their CPUs made at the same gigantic foundry, known as the Taiwan Semiconductor Manufacturing Company, or TSMC for short. So much of the world’s leading-edge semiconductor production runs through TSMC that its operation has become a geopolitical issue. However, in America, Intel has fab facilities scattered across the West, in addition to TSMC’s Arizona foundry. Intel also has facilities in Ireland (Leixlip) and Israel (Kiryat Gat). In 2009, AMD spun off its manufacturing arm into GlobalFoundries, which has a fab in Dresden and another, massive fab (plus its headquarters) in Malta, New York.