Clicked Gallery
What is a GPU (graphics processing unit)?
Highlighted from a real engineering doc. Explained by Clicked.
Used in a sentence
Engineering Notes · AI Systems
The startup had spent most of its new funding renting GPU clusters to train its models.
The reader highlighted one word in the docs. Clicked explained the technical term “GPU” in simple terms:
Explained in three depths
●○○
Overview
A GPU, short for graphics processing unit, is a computer chip that applies one simple operation to thousands of numbers at once. A CPU works in depth, taking one task through many varied steps. A GPU works in width, taking one step and running it across all the data at the same time. Drawing a screen is width work, which is why the chip was built for graphics. Training AI is width work at a far larger scale, which is why the same chip now sits at the center of AI.
●●○
Detail
A GPU is a chip that speeds computers up by doing thousands of things at once. A CPU has a few powerful cores taking a task through varied steps in order; a GPU has thousands of simple cores all performing the same step at once, each on its own piece of the data. GPUs got this design because of what screens demand. An image is millions of pixels that all need the same arithmetic every frame. One core taking them in turn could never keep up, so graphics chips grew thousands of cores to handle the batch together. Decades later, AI training turned out to be the same shape of work. A neural network is mostly multiplication over huge grids of numbers, the same operation repeated billions of times, so the graphics chip was already the right tool. The design matters more every year. Individual cores stopped getting much faster years ago, so further speed has to come from doing more at once. The limit: only work that splits into identical, independent pieces can be shared out this way. A job where every step waits on the last, a tax form worked line by line, stays on the CPU no matter how many GPU cores sit idle. Training a large model is the extreme case: thousands of data-center GPUs stay busy for months, and when the news says a lab has "compute", GPUs are what it is counting.
●●●
Analogy
Grading exams shows the split. A thousand multiple-choice sheets need the same answer key applied a thousand times. One professor checking sheet after sheet takes all night. Five hundred helpers, each given the answer key and two sheets, finish in minutes. The helpers are not skilled; the answer key carries all the judgment, and that is the GPU: many simple hands, one shared instruction. The essay is the other kind of work. It is one argument that must be followed start to finish, so it goes to the professor alone, and that is the CPU. A computer splits its work the same way: the thousand identical checks go to the GPU, and the one argument that must be read in order goes to the CPU.
AI explanations may contain errors · Not professional advice
Same facts, different vibe — Slang mode 😎
●○○
Overview
A GPU is the chip the whole AI industry is queuing for, and it was built to make video games look good. The trick never changed: the same tiny calculation, everywhere, all at once. Rendering an explosion means running that calculation across a million pixels at once, so gamers funded twenty years of its development. Then chatbots turned out to need the exact same trick at a thousand times the dose, and the gaming chip found itself running the AI boom. 😎
●●○
Detail
A GPU does an absurd number of small calculations at once, and that one ability now moves markets. Read any AI story and the word "compute" appears. Compute means GPUs. Startups raise fortunes and spend them renting these chips by the hour. A single data-center unit costs more than a decent car, whole clusters run to tens of thousands of units, and companies quote their GPU counts the way navies quote ships. All of this for a part whose name still says graphics processing unit. The joke is that the AI versions never touch graphics. They sit racked in data centers doing pure math with no screen attached, and the name simply never got updated. The one in your laptop still does the original job, pushing pixels to your display. And one catch stays out of the headlines: a GPU only earns its price when the work splits into identical pieces and the data arrives fast enough to keep it fed. Get either wrong and you have bought the world's most expensive way to wait. 😎
●●●
Analogy
A GPU is apple-picking day done with an army of volunteers, one per tree. The star picker, the CPU of this story, clears a tree faster than anyone alive, and left to himself he would need a week. The volunteers get one rule, pick everything red, and the orchard is bare by lunch. Not one volunteer is half the picker the star is. The speed came from dividing the job, not from talent. The fine print lives there too. Bussing in a crowd that size needs organizing, and if the crates show up late, hundreds of pickers stand around watching one truck. And the grafting, where every cut depends on the cut before, is no job for a crowd. That kind of task goes back to the star picker, the CPU, which is why every laptop still ships with one. 😎
AI explanations may contain errors · Not professional advice
Formal definition — The same term, explained the usual way
A graphics processing unit is a specialized processor containing thousands of arithmetic cores that execute the same instruction across large blocks of data in parallel. Originally developed to render real-time graphics, the architecture proved efficient for any workload dominated by dense linear algebra, including scientific simulation and the training and inference of neural networks. It complements rather than replaces the central processing unit, which retains sequential and control-flow work; data-center variants add high-bandwidth memory and high-speed interconnects to operate in clusters of thousands of chips.
Want Clicked to explain terms like “GPU” directly in your browser — including on PDFs?
Add to Chrome — Free50 free Explanations · No credit card required
More from the gallery
What is a Neural Network?
Math that learns the rules from examples instead of having them written, and why that wins.
What is training (in AI)?
The process that builds a model by example. Chatting with it afterwards changes nothing.
What is Inference (in AI)?
The model doing its actual job, and why the cheap half ends up costing the most.
What is Quantization?
Storing a model's numbers less precisely — half the memory, barely any loss.
What is a large language model (LLM)?
A program that predicts what text comes next, then taught separately to answer rather than ramble.
What is Edge Computing?
Doing the work on the device instead of a data centre, and the three situations that require it.