Short answer
A GPU (graphics processing unit) is a processor built to perform thousands of simple calculations at the same time. It was originally designed to draw 3D graphics for games, which means calculating millions of pixels in parallel. That same ability makes GPUs ideal for artificial intelligence, video editing and scientific computing. A CPU has a handful of powerful cores for general tasks; a GPU has thousands of smaller cores plus its own fast memory (VRAM), and modern GPUs add tensor cores built specifically for AI math.
What does a GPU do?
Every frame you see in a game is made of millions of pixels, and each one needs its colour, lighting and position calculated, 60 or more times per second. Those calculations are similar and independent, so they can be done side by side. A GPU is built for exactly this kind of parallel work.
Today GPUs are used for:
- Gaming and 3D graphics: rendering scenes, lighting, ray tracing and upscaling such as DLSS.
- Artificial intelligence: training and running neural networks, from chatbots to image generators.
- Video: editing, effects, encoding and streaming.
- Science and engineering: simulations of weather, molecules, fluids and physics.
- Data analysis: speeding up large database queries and data science workloads.
- Displaying your screen: every computer needs some form of GPU to output an image.
GPU vs CPU
| CPU | GPU | |
|---|---|---|
| Cores | A few to a few dozen powerful cores (up to ~192 on servers) | Thousands of smaller cores (21,760 CUDA cores on an RTX 5090) |
| Best at | Varied tasks, complex logic, running the operating system | The same operation on huge amounts of data at once |
| Analogy | A few expert chefs who can cook anything | Thousands of kitchen helpers each chopping vegetables |
| Memory | System RAM, large but slower | VRAM, smaller but much faster bandwidth |
| AI use | Small models, data preparation | Training and running most AI models |
The CPU and GPU work together: the CPU runs the program and hands the heavy parallel maths to the GPU. For a deeper comparison, read GPU vs CPU for AI.
Integrated vs dedicated GPUs
- Integrated GPU: built into the processor chip and shares system memory. Found in most laptops and office PCs, and in phones. Fine for everyday use, video and light games. Apple Silicon Macs use a powerful integrated GPU with large shared "unified memory".
- Dedicated (discrete) GPU: a separate graphics card or chip with its own VRAM and cooling. Needed for serious gaming, 3D work and most local AI.
- Data center GPUs: chips like the NVIDIA H100, H200 and B200 or AMD Instinct MI300X, built for AI and supercomputers, with huge memory and no display outputs.
What is inside a GPU?
- Streaming multiprocessors (SMs): the building blocks that hold the cores. Explained in how a GPU works.
- CUDA cores (NVIDIA) or stream processors (AMD): general arithmetic units. See What are CUDA cores?.
- Tensor cores: specialised units for the matrix multiplications at the heart of AI. See tensor cores vs CUDA cores.
- RT cores: hardware for ray-traced lighting in games. See tensor cores vs RT cores.
- VRAM: the GPU's own fast memory (GDDR7 on consumer cards, HBM on data center chips). See What is VRAM?.
- Memory bus and bandwidth: how fast data flows between VRAM and the cores. See GPU memory bandwidth.
Why AI runs on GPUs
Neural networks are mostly giant matrix multiplications: multiplying large grids of numbers together, billions of times. That work is massively parallel, just like graphics. Three moments made GPUs the engine of the AI boom:
1. 2007: CUDA. NVIDIA released CUDA, letting programmers use GPUs for general computing, not just graphics. See What is CUDA?. 2. 2012: AlexNet. A neural network trained on two NVIDIA GTX 580 gaming GPUs won the ImageNet image-recognition contest by a wide margin, kick-starting the deep learning era. 3. 2017: tensor cores. NVIDIA's Volta V100 added tensor cores dedicated to AI maths, multiplying AI performance.
Today's large language models are trained on clusters of thousands of data center GPUs (see What is a GPU cluster?), and every ChatGPT-style answer is generated on GPUs or similar AI chips.
Who makes GPUs?
- NVIDIA: GeForce for gaming, RTX PRO for workstations, and H100/H200/B200-class chips that dominate AI data centers.
- AMD: Radeon for gaming and Instinct accelerators for AI and supercomputing. See AMD vs NVIDIA for AI.
- Intel: Arc graphics cards and integrated graphics.
- Apple, Qualcomm and Arm-based chipmakers: integrated GPUs in Macs, phones and tablets.
Google's TPUs and other AI accelerators are not GPUs, but they compete with GPUs for AI work. See NPU vs GPU vs TPU.
How to choose a GPU
- For everyday use and office work: an integrated GPU is enough.
- For gaming: match the GPU to your monitor resolution and refresh rate.
- For video editing and 3D: more VRAM and strong encoders help.
- For AI: VRAM decides which models you can run; memory bandwidth decides how fast. 16 GB is a good minimum for local AI. See best GPU for local LLMs and best GPU for Stable Diffusion.
- For big AI jobs: renting is often cheaper than buying; compare today's prices in our GPU cloud price tracker.
To compare models in detail, learn how to read a GPU spec sheet.
Frequently asked questions
Is a GPU the same as a graphics card?
Not quite. The GPU is the chip itself. A graphics card is the whole board that holds the GPU, its VRAM, power delivery, cooling and display outputs.
Does my computer have a GPU?
Yes. Every computer that shows an image has a GPU, either integrated into the processor or as a separate card. On Windows, open Task Manager and look under Performance to see it.
Why are GPUs so expensive?
Demand from AI data centers, gaming and crypto has repeatedly outpaced supply, and high-end GPUs use advanced, costly chip manufacturing and memory. See why GPU prices are rising.
What was the first GPU?
NVIDIA marketed its GeForce 256, released in 1999, as the world's first GPU, because it moved 3D transform and lighting calculations from the CPU onto the graphics chip.
Sources and further reading
- NVIDIA, GeForce RTX 5090 specifications
- NVIDIA, CUDA Zone
- Krizhevsky, Sutskever and Hinton, ImageNet Classification with Deep Convolutional Neural Networks (AlexNet, 2012)
- NVIDIA, NVIDIA Tesla V100 GPU Architecture whitepaper (2017)




Leave a Reply