The History of AI

TPU

Era 4 · Deep learning and foundation models (2012–2021) › Infrastructure and institutions

Lineage
Machines and computation
Kind
Implementation

Who

Google

What happened

At its developer conference, Google I/O, on May 18, 2016, Google unveils the tensor processing unit (TPU), a chip designed to run neural networks that it has used in its data centers since 2015.

Why it mattered

The first chip designed for neural networks to be deployed at scale. It opens the race for specialized AI hardware.

The full story

When Google began using neural networks in services such as search, photos, and voice, the cost of running them for millions of users became a problem for its data centers. Rather than rely only on general-purpose processors, the company designed its own chip for the operation that takes the most time in a neural network: multiplying matrices.

The first TPU, whose lead architect was Norman Jouppi, was a matrix multiplication engine working with 8-bit numbers, a low precision that is enough for running an already trained network. Its core had 65,536 multiply-accumulate units and reached 92 trillion operations per second.

Google had been using it in its data centers since 2015, for search with RankBrain, Google Photos, and Street View, and also in AlphaGo's games against Lee Sedol. It announced the chip on May 18, 2016. In the 2017 technical paper, the company reported that it was on average 15 to 30 times faster than a contemporary CPU and GPU and 30 to 80 times more energy efficient.

The second generation, from 2017, could also train networks, and since 2018 Google has rented TPUs out to others. The TPU started the race for specialized AI hardware.

Leads to

  • AlphaGo Zero and AlphaZero (October 19, 2017)

    Documented direct influence

    AlphaGo Zero plays on four TPUs, the chip Google had unveiled in May 2016

Sources

How to cite

“TPU.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/tpu

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline