TPU

What is a TPU?

A TPU — tensor processing unit — is a custom chip Google designed specifically to run machine learning. If a GPU is a generalist who picked up AI later, a TPU is a specialist that was born for it: the architecture is built around tensor operations, the math underneath AI, from day one.

How's it different from a GPU?

Different mission
A GPU is a general parallel-computing chip — games, rendering, AI, all of it. A TPU is a deeply customized AI accelerator and nothing else.
Different access
You mostly rent TPUs as a cloud service through Google Cloud. You're not buying one to slot into your desktop.
Different philosophy
TPUs trade away generality to win on efficiency and throughput for specific AI workloads.

Why does it matter?

Google's own models — including the Gemini family — are trained on massive TPU clusters. TPUs proved the "custom chip" path works: lower power, higher AI efficiency, and it means Google doesn't have to lean entirely on NVIDIA for compute.

Will I ever touch a TPU?

Probably already have, without knowing. Google Search, Translate, YouTube recommendations — there may well be TPUs doing the quiet work behind them. Developers can even grab a slice of TPU for free in Google Colab to run experiments.

Bottom line: a TPU is Google's purpose-built chip for AI, proof that a dedicated specialist can hold its own.