Definition
AI hardware is the set of processors and memory that run models. A CPU is a general-purpose processor with a few powerful cores. A GPU has thousands of simpler cores that work in parallel. TPUs and NPUs are chips built specifically for AI maths, in data centres and on devices respectively.
In simple words
A CPU is a brilliant professor who solves one hard problem at a time. A GPU is a hall of ten thousand students, each doing one small sum at the same moment. Training a network means billions of small sums, so the hall wins easily.
Why it matters
Hardware decides what you can train, how long it takes and what it costs. It's also why GPU makers became some of the world's most valuable companies. You don't need your own: Colab and cloud GPUs are enough to start.
How it works
| Chip | Good at | Used for |
|---|---|---|
| CPU | Varied work, one step after another | Classic ML, data prep, running apps |
| GPU | The same maths on many numbers at once | Deep learning training and inference |
| TPU | Tensor maths, at data-centre scale | Google's AI workloads |
| NPU | AI tasks using little power | Phones and laptops |
What to look at
- GPU memory (VRAM) often matters most: the model must fit, or it won't run at all.
- Moving data can slow everything down, so storage and memory speed matter as well.
- Cost and electricity are real limits on what you can run.
Common mistakes
- Buying an expensive GPU to start learning. Free notebooks with GPUs are enough for beginners.
- Using a GPU for small classic ML jobs that run fine on a CPU.
Check yourself
Summary
- CPU: a few strong cores for varied work. GPU: thousands of cores for parallel maths.
- TPUs and NPUs are specialised AI chips.
- VRAM usually sets what you can run.
Finished reading?