Tech, ExplainedSemiconductors and devices
Concept

CPU, GPU, SoC and FPGA

Four ways to build a processor.

HOW IT WORKSA program haswork to do: variedsteps, or one sum onmillions of numbersVaried steps go, oneafter another, to theCPU's few large coresOne repeated sumis split acrossthe GPU's thousandsof small coresIn a phone, the CPU,GPU and modem shareone SoC to save spaceand powerFor one special job,an FPGA is rewired byloading a new fileEach kind of processor gives up some flexibility to be faster at one sort of work.HOW IT WORKSA program haswork to do:varied steps,or one sumon millionsof numbersVaried stepsgo, one afteranother, tothe CPU's fewlarge coresOne repeatedsum is splitacross theGPU'sthousands ofsmall coresIn a phone,the CPU, GPUand modemshare one SoCto save spaceand powerFor onespecial job,an FPGA isrewired byloading anew fileEach kind of processor gives up some flexibility to be faster at one sort of work.HOW IT WORKSA program has work to do: varied steps,or one sum on millions of numbers↓Varied steps go, one after another, tothe CPU's few large cores↓One repeated sum is split across theGPU's thousands of small cores↓In a phone, the CPU, GPU and modem shareone SoC to save space and power↓For one special job, an FPGA is rewiredby loading a new fileEach kind of processor gives up some flexibility to befaster at one sort of work.
Varied tasks go one after another to a CPU's few large cores, one repeated sum goes to a GPU's thousands of small cores, a phone holds both on one SoC, and an FPGA is rewired by loading a new file.
opens a page inside this one a page in the same scene a page in another scene a concept page a label only, with no page

CPU, GPU, SoC and FPGA are four kinds of processor, the chip that computes. Each is built for a different sort of work. A CPU, the central processing unit, does many different tasks, one after another, very fast. A GPU, the graphics processing unit, does the same sum on thousands of numbers at once. An SoC, a system-on-chip, puts a CPU, a GPU and other parts together on one chip. An FPGA, a field-programmable gate array, has circuits that can be rewired after it is made.

All four are built from transistors, tiny electrical switches, with billions of them on one chip. The transistors are grouped into cores. A core is the part of a processor that reads instructions and carries them out.

Here is how the four differ. A CPU has a few large cores, from about 8 in a laptop to more than 100 in a large server. Each core is built for varied steps: read a file, compare two values, choose what to do next. It runs the operating system and the apps. A GPU has thousands of small, simple cores. It was built to work out the colour of millions of pixels at once, which is the same sum repeated. Training and running AI models needs exactly that kind of arithmetic: multiplying large tables of numbers. So GPUs now do most AI work. An SoC holds a whole system on one piece of silicon: a CPU, a GPU, an AI engine, the modem that connects to the mobile network, and more. Putting them together saves space and battery power, so every smartphone runs on a system-on-chip. An FPGA is a grid of small logic blocks whose connections are set by a file loaded each time the chip is switched on. Load a different file, and the same chip becomes a different circuit. It is slower and costlier per chip than a fixed design, but it can be changed.

Take, for example, a train-booking website at 10 a.m., when Tatkal booking opens. Its servers' CPUs handle each request in turn: check the seat, take the payment, issue the ticket.

Consider an AI lab training a language model. It uses thousands of GPUs, because the training is almost entirely large multiplications of numbers.

Say a team is designing a new chip. Before the design is manufactured, the team loads it onto FPGAs, or onto a hardware emulator, and runs real software on it to find mistakes.

How do these four differ from an ASIC, an application-specific integrated circuit? An ASIC is built for one job only, and cannot be rewired. It is the fastest and most efficient option for that job. But designing one costs hundreds or thousands of crores of rupees, so it pays off only when it is used in very large numbers. A custom AI ASIC, built for one cloud company's work, is one example. AI accelerators in general are built on the same idea.

Appears in

Semiconductors and devices
Tech, Explained is a part of ingrid.fyi. Every page explains one piece of technology in plain words.