Picture a hyper-specialised worker who can do only one thing — multiply and add arrays of numbers — but does it thousands of times faster than any generalist. That is the idea behind the TPU (Tensor Processing Unit), a chip designed by Google in 2015 to accelerate machine learning. It is an ASIC: a circuit custom-built for a single purpose, unlike the all-purpose CPU.
Why "tensor"?
In AI, almost everything reduces to matrix multiplications. A neural network chains operations of the form:
$$ Y = X \cdot W + b $$
where $X$ is the input, $W$ the weights and $b$ the bias. The TPU uses a systolic array: a grid of thousands of tiny compute cells that pass data hand-to-hand, without endlessly fetching from memory. This cuts the energy bottleneck and massively speeds up processing.
TPU vs GPU vs CPU
| Criterion | CPU | GPU | TPU |
|---|---|---|---|
| Purpose | General | Graphics / parallel | AI (matrices) |
| Flexibility | Very high | High | Low (specialised) |
| AI efficiency | Low | Good | Excellent |
The CPU is a Swiss Army knife, the GPU a parallel assembly line, the TPU a dedicated machine tool.
Real-world use
TPUs power Google Search, Google Translate and the training of large models such as Gemini. Developers can access them through the cloud (Google Cloud TPU).
When a computation is repeated billions of times, etching it into silicon becomes the decisive lever for performance.