Products, laws and numbers here change fast. Checked October 2026.
In plain words
Every computer has a main chip, the CPU. It has a few very clever cores that do tasks one after another, fast. It is great at running your apps.
A GPU, or graphics processing unit, was built for video games. To draw a game, a computer must colour millions of dots on the screen, many times a second. So a GPU has thousands of simpler cores that each handle a small piece, all at the same time.
AI turns out to need the same thing. Training a model means doing the same small sum (multiply, then add) billions of times over. A GPU spreads those sums across all its cores at once. That is why large language models are trained on thousands of GPUs linked together in a data center.
Everyday example
Checking 10,000 exam papers
A board has 10,000 answer sheets to mark. One brilliant professor could do them all, but it would take months. Instead, 1,000 teachers each take ten sheets and finish in an afternoon. Each teacher is less brilliant than the professor, but the job is simple and repetitive, so numbers win. A CPU is the professor. A GPU is the room full of teachers.
Where the example breaks: Teachers can each think for themselves. GPU cores are much simpler: they do the same instruction on different numbers, in lockstep, and need the CPU to hand out the work.
How it actually works
AI work is mostly one operation: multiplying large grids of numbers, called matrices. GPUs are built to do that in parallel.
Many cores, one instruction
A GPU groups its cores so that each group runs the same instruction on different numbers at once. Drawing a frame or running a neural network layer both fit this pattern: lots of independent sums with the same shape.
From graphics to general maths
In November 2006 Nvidia introduced CUDA, software that let programmers use the GPU for any calculation, not only graphics. Researchers began using it for science, and then for neural networks.
The 2012 moment
AlexNet, an image-recognition network, was trained in five to six days on two consumer GPUs and won the ImageNet contest by a wide margin. After that, deep learning and GPUs grew up together.
Fast memory matters as much as cores
A model's parameters must stream into the cores constantly. Data center GPUs sit next to stacks of very fast memory (HBM) so the cores are not left waiting.
- Name first used for
- Nvidia's GeForce 256 graphics card, released October 1999 and marketed as the world's first GPU
- Turned general purpose
- CUDA, November 2006
- Scale of one chip
- Nvidia's Blackwell B200 (announced March 2024) has 208 billion transistors and 192 GB of fast memory
- As of October 2026
- Nvidia's Blackwell chips are widely used; its next generation, Rubin, reached first customers as samples in early 2026, with volume shipments planned for the second half of the year. AMD and others also make AI GPUs
GPUs are not the only AI chips. Google's TPU and the NPU in your phone are built only for AI maths.
Questions people ask
What is the difference between a CPU and a GPU?
A CPU has a few powerful cores that handle varied tasks one after another very quickly. A GPU has hundreds to thousands of simpler cores that do the same kind of sum on lots of data at once. Your computer uses both: the CPU runs things, the GPU handles heavy parallel work.
Why are GPUs used for AI?
Training and running AI models is mostly multiplying huge grids of numbers. Those sums are independent of each other, so they can all happen at once. A GPU's thousands of cores do exactly that, many times faster than a CPU for this kind of work.
Do you need a GPU to use ChatGPT?
No. When you use ChatGPT or a similar app, the model runs on GPUs in the company's data centers and only the answer comes to your phone. You need a strong GPU only if you want to run a large open-weight model on your own computer.
Why is Nvidia so important to AI?
Is a graphics card the same as a GPU?
Almost. The GPU is the chip itself. A graphics card is the board that carries the GPU, its memory and its cooling fans, and plugs into a computer. Data center GPUs come on boards built for servers instead, often without any screen output.
Further reading
- What's the Difference Between a CPU and a GPU?Nvidia, 2009
- ImageNet Classification with Deep Convolutional Neural Networks (Krizhevsky, Sutskever, Hinton)NeurIPS, 2012
- Nvidia introduces CUDA architecture for computingComputer Graphics World, 2006
- Nvidia's Blackwell B200 GPU revealedTom's Hardware, 2024
- First Vera Rubin AI chips are rolling out to customersTechRadar, 2026