Anniversaries, 18 May 2027
Google TPU revealed (2016)
On 18 May 2016 Google revealed the Tensor Processing Unit, a chip it had built to run neural networks in its data centres.
At its I/O developer conference, Google disclosed that it had been using a custom AI chip in its data centres for over a year. The first TPU was built around a systolic array of 65,536 8-bit multiply-accumulate units and was designed for running trained models, a task called inference. Later generations added training and were linked in large pods. The TPU showed that a cloud company could design its own accelerator, and others, including Amazon, Microsoft and Meta, followed.