GPU
in English: graphics processing unit (GPU)
A GPU is a computer chip that performs many simple calculations simultaneously, making it today's most important tool for training and running artificial intelligence.
In plain terms
The GPU originally rendered images for video games: it had to calculate the color of millions of pixels on the screen many times per second. This required a chip that does not solve a single complex task quickly, but many simple ones in parallel.
Imagine it this way: your computer's traditional processor (the CPU) is a few excellent chefs who can prepare any dish but only a few at a time. The GPU, by contrast, is a kitchen brigade of several thousand people where everyone only peels potatoes, but all at once. If you need ten thousand servings of potatoes, the brigade wins.
Artificial intelligence is precisely this kind of 'potato-peeling' task: the operation of language models is essentially massive amounts of multiplication and addition. This is why the GPU became the most important component of the AI era, and why news reports focus on which company or country can purchase how many units.
How it works
- The chip contains thousands of small computing units (cores) that perform the same operation simultaneously on different data.
- The program breaks the large task into many small, independent parts. For AI models, these are largely multiplications of huge number tables (matrices).
- Data moves into the GPU's own fast memory (VRAM). Its size determines how large a model fits on a card.
- The cores calculate partial results in parallel, and the system assembles them. This produces a frame of an image or the next word, or more precisely, word fragment, of a chatbot's response.
- In data centers, many GPUs are connected, with tens or even hundreds of thousands in the largest systems. Together, they train a large model in weeks or months.
An everyday example
When you ask a question to a chatbot, the answer is usually not generated on your phone but in a remote data center, where GPUs or other AI chips calculate it word by word. If you read in a laptop advertisement that it has a 'dedicated GPU' or 'dedicated graphics card', it means that in addition to the graphics unit built into the processor, there is a separate GPU inside. This allows games, video editing, and smaller AI models to run faster on it.
Who is affected
A common misconception
Limits and risks
- Expensive and sometimes scarce: top-tier AI chip prices rival those of a car, and supply is concentrated in the hands of a few manufacturers.
- It consumes a lot of electricity and runs very hot, so the energy and cooling needs of large data centers are a serious environmental and power-grid issue.
- Memory size is a hard limit: if the model does not fit in GPU memory, it runs slowly or not at all.
- Multiple GPUs do not automatically mean smarter AI: model quality also depends on data and design.
Similar, but not the same
- CPU (central processing unit) – The CPU has few but versatile cores that solve any task one after another, while the GPU has many simple cores that perform the same operation massively and simultaneously.
- graphics card – The graphics card is the entire component with cooling, memory, and connectors, while the GPU is the computing chip mounted on it, although common usage often conflates the two.
- NPU (neural processor) – An NPU is an energy-efficient chip designed specifically for AI tasks in phones and laptops. A GPU is more general-purpose, usually more powerful, but consumes more power.
Related concepts
Hardver, computing capacity, AI infrastructure, AI accelerator
CPU, NPU, graphics card, Data center, Model training, Inference
A brief history
The term 'GPU' was popularized by Nvidia in 1999 with the launch of the GeForce 256. In 2007, Nvidia released its CUDA development environment, making it easy to use GPUs for calculations beyond graphics. The turning point came in 2012: the AlexNet image recognition neural network was trained on two GPUs intended for gaming and won the ImageNet image recognition competition by a large margin. Since then, deep learning has occurred predominantly on GPUs, alongside other AI accelerators such as Google TPUs.
Where things stand
At the end of 2026, the GPU is the basic resource of the AI industry. Training and serving large models, as well as running AI agents, largely takes place in GPU-based data centers, although the large cloud companies' own AI chips are also becoming more widespread. Demand for top-end chips remains very strong. Access has also become an economic and geopolitical issue, for example because of US export restrictions.
In the news
- AI data center construction shifts to a manufacturing mode: faster build-out, more flexibilityOct 28, 2026 · Ipar
- Power and skilled labour, plus cooling and GPU links, constrain AI data centre buildsOct 12, 2026 · Ipar
- AI writes GPU code that even CUDA engineers do not fully understandOct 2, 2026 · Code
- AI GPU cluster responds to ERCOT peak-demand signal within a secondAug 24, 2026 · Ipar
- Nvidia lines up $500 billion with Wall Street partners to finance AI infrastructureAug 10, 2026 · Agent Age
Also known as
GPU · grafikus processzor · graphics processing unit · graphics card · graphics card · graphics processing unit · graphics card
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.