What are TPUs and NPUs?
TPUs and NPUs are chips or chip parts built just for AI math, so they run AI faster and with less energy than a CPU.
1. AI is mostly grid math
Most AI work is multiplying big grids of numbers, called matrices. One reply from an AI chat app can take billions of small multiplications.
2. Built for one kind of math
TPUs and NPUs are like a pasta machine beside a cook: built mainly for grid math, they do it fast with little energy. They still run many different AI programs, because they all use grid math.
3. TPUs fill data centers
TPU, short for tensor processing unit, is Google's name for its own AI chips. Thousands work together in data centers, buildings full of computers, to train AI (teach it from examples) and run it.
4. NPUs in laptops and phones
An NPU, or neural processing unit, is the AI engine inside many laptop and phone processors. It runs features such as background blur in video calls.
5. Speed levels and sleep
The NPU has its own speed levels, so it runs only as fast as the work needs. It runs slower for lighter work or when the chip is hot, and sleeps when there is no AI work.
Words to know
- Matrix (plural: matrices)
- A grid of numbers, in rows and columns.
- AI (artificial intelligence)
- Programs that learn patterns from many examples.
- TPU (tensor processing unit)
- Google's name for its own AI chips, used in its data centers.
- Tensor
- Numbers set out in a grid, or grids stacked into a block.
- NPU (neural processing unit)
- The AI engine inside many laptop and phone processors.
- Data center
- A building full of computers that serve many users at once.
Where you meet it at work
Power tests run AI test programs and check that the NPU speeds up when work arrives, then sleeps when it ends.
Check yourself
Why does an NPU use less energy for AI than a CPU?Think first, then tap to see the answer.
It is built for grid math, so it does that math with less waste.