Clicked Gallery

What is CUDA?

Highlighted from a real engineering doc. Explained by Clicked.

Used in a sentence

Engineering Notes · AI Systems

The lab's entire training pipeline was written in CUDA, which made the choice of hardware automatic.

The reader highlighted one word in the docs. Clicked explained the technical term “CUDA” in plain language:

Explained in three depths

Same facts, different vibe — Slang mode 😎

The Clicked way

●○○

Overview

CUDA is the software through which programmers use Nvidia's graphics chips for general computing, anything from physics simulations to training AI models, rather than for drawing graphics. A programmer writes in an ordinary language such as C++ or Python, and CUDA turns that program into work the chip's thousands of cores can carry out together. Nvidia shipped it in 2007; the first card able to run it, from 2006, carried 128 cores. It does not make the chip faster. It makes the chip reachable.
●○○

Overview

CUDA is why a gaming rig can moonlight as a lab supercomputer. The silicon was born to paint pixels, and the same talent multiplies numbers beautifully, if anything can ferry the math in. CUDA is the ferry. Write 40 plain-looking lines, and the machinery turns them into 4,000 little jobs happening simultaneously. The maker gives the whole kit away free, which stops looking like charity when you notice the fine print: everything anyone builds with it lands on one brand of silicon, and moving house later means rebuilding, not repacking. 😎

A quick take — often all you need.

●●○

Detail

CUDA is Nvidia's platform for programming its graphics chips: the language additions, tools and ready-made libraries through which most software reaches a GPU's thousands of cores. A GPU takes its work in a particular shape, and before 2007 the practical way in was graphics software, so other problems had to be rewritten as graphics work. CUDA gave general programs a direct way in, and the field's code, libraries and trained people have accumulated on it since. Two things are easy to confuse here. The chip is open: other companies write their own software for talking to these GPUs, and it works. The code is not: a program written on CUDA runs on Nvidia's chips and, in practice, nowhere else. Moving it would mean rebuilding CUDA's libraries for another maker's chips, an enormous amount of software that changes with every release and belongs to Nvidia. Attempts exist and stay years behind. A software layer usually promises that the thing underneath can be swapped, and that is exactly the promise CUDA does not make. So new AI work goes where the existing code already runs, and picking the software has picked the hardware. Of the world's 500 fastest supercomputers in November 2024, more than a third ran on Nvidia's GPUs.
●●○

Detail

CUDA is the half of the AI story that compounds: the software. Say your lab has 10 years of in-house programs, 300,000 lines, all of them written against CUDA, because that is what everyone wrote against. Now a competitor's chip turns up, genuinely decent. The sticker price is fine; the real price is the rewrite: 300,000 lines re-plumbed, each library traded for a younger clone, everyone retrained, and the old machines still humming in the corner the whole time. And plenty of people sell escape routes, translation layers built for exactly this problem, and the labs still mostly stay. Illustrative numbers, real arithmetic, and it is the arithmetic a lab does before deciding to stay put. Which is why the fight in AI silicon is not only about who makes the quickest chip. It is also about the mountain of code the world has written, and the mountain grows daily, in one direction. 😎

Want more? One click digs deeper.

●●●

Analogy

CUDA is the score, and the GPU is the orchestra. Its 100 players sit ready, and no composer walks the hall telling each one what to do; that would take all night and ruin the piece. The score is the form that speaks to all of them at once: written once, read in parallel, every player finding their own line. Learning to write for an orchestra means learning to write in that form. Hand the same hundred players an essay instead and nothing happens, however brilliant the essay: the players are willing, but the form is wrong. That is CUDA: not the musicians, not the music, but the form the work must take to reach a hundred hands at once.
Deepest level reached.
●●●

Analogy

CUDA is the battery system your power tools all share. The first cordless drill you buy quietly decides the next ten purchases, because the saw, the sander and the strimmer all snap onto the same battery, and none of them fits the other brand's tools. The batteries store the same electricity. The tools do the same jobs. It does not matter: by the fifth tool you are not choosing tools anymore, you are choosing to stay. Swapping brands means repurchasing the lot, so almost nobody does. That is CUDA: every program written for it is another tool on the same battery. 😎
Deepest level reached.

Unfamiliar concept? A real-world example makes it click — fresh analogies on tap.

AI explanations may contain errors · Not professional advice

Formal definition — The same term, explained the usual way

CUDA (Compute Unified Device Architecture) is a parallel-computing platform and programming model introduced by Nvidia in 2007, comprising compiler support, language extensions for C, C++ and other languages, runtime software and an extensive set of accelerated libraries. It allows general-purpose programs to execute across the many cores of Nvidia GPUs, a practice earlier possible mainly through graphics interfaces. CUDA is proprietary and runs on Nvidia hardware only; open alternatives such as OpenCL exist, and other manufacturers ship their own programming layers, though CUDA's accumulated libraries and installed base remain the largest. The platform is free to use; the hardware it addresses is not.

Want Clicked to explain terms like “CUDA” directly in your browser — including on PDFs?

Add to Chrome — Free

50 free Explanations · No credit card required