Explain like I'm five · Systems

What is GPU/GPGPU?

GPU and GPGPU, told as a story about a crew of helpers doing the same simple task all at once.

hundreds of cars, each needing the exact same tire checkthe chief (CPU) has few hands, so only a few get checked at a time

The same tire check piles up by the hundreds.The chief (CPU) has few hands, so only a few get checked at a time.

I'll do it faster, better!even a far better single mechanic still handles one car at a time

Even a far better single mechanic still checks one car at a time.Clearing the piled-up checks fast takes many hands, not one better one.

TIRE CHECKGPU = a crew of hundreds doing the same simple task all at onceeach helper does only simple work, but together they finish the whole check far sooner

GPU = a crew of hundreds doing the same simple task all at once.Each helper does only simple work, but together they finish the whole check far sooner.

Hundreds–Thousands of Coresthat many hands working at once
SIMDthe same move, given to everyone at once
Throughput-Firstfinishes many simple tasks, rather than one complex task
GPGPUthe graphics crew, repurposed for general math
CPU vs GPUCPU: a few skilled mechanicsCPU: complex, varied jobsGPU: hundreds-thousands of helpersGPU: the same job, all at onceGPU shines at many simple tasks at once, not one complex task

CPU and GPU are good at different kinds of work.GPU shines at many simple tasks at once.

the same task, hundreds of timesGPU beats the CPU (chief) by a landslideone complex job full of branching decisionssuch work is still the chief's (CPU's) jobGPU doesn't replace the CPU — they split different kinds of work

GPU isn't suited to complex work full of branching decisions.GPU doesn't replace the CPU — they split different kinds of work.

In one breath

GPU = a crew of hundreds to thousands of small hands doing the same simple task all at once. GPGPU is putting that crew to work on general computation, not just graphics.

A GPU (Graphics Processing Unit) applies the same instruction to many data items at once across thousands of simple cores — a SIMD (Single Instruction, Multiple Data) design. Built originally for parallel pixel and vertex math, it also took on non-graphics calculations, which gave rise to the term GPGPU (General-Purpose computing on GPU); today matrix-heavy deep learning is its best-known use. A CPU excels at fast, branching, sequential logic with few powerful cores; a GPU excels at maximizing throughput with many simple ones.

When grown-ups say it

SIMD
Giving everyone the same move at once. The core pattern behind GPU parallelism.
GPGPU
Putting the graphics crew to work on general math. Deep-learning training is the classic example.
Throughput
How much gets processed in a given time. This is what GPU design optimizes for.
AI Accelerator (TPU/NPU)
A robot arm built for exactly one calculation. An even more specialized step beyond the GPU. → the robot arm built for one calculation
CPU Scheduling
The rule deciding who gets the lift first. How CPU turns get decided; on that page the chief is the OS, not the CPU. → who gets the lift first
Process/Thread
Several hands working the same car. Both split work among many hands, but a GPU specializes in running the same calculation at once. → one car, many hands
Time Complexity
The ruler for how fast time grows as work piles up. Splitting work in parallel changes the felt speed even at the same Big-O. → how much slower as cars pile up
GPU in the AI Field
The fast hardware (perch) that runs an AI model (the parrot). The AI field (the AI field) covers it briefly, from the inference-cost angle. → AI inference
Next story
Boxes Get Stacked, Cars Get Queued →
Want another term as a picture book? Request →