What is the name of Google’s custom AI chip first announced in 2016?

Google’s custom AI chip first announced in 2016 is the Tensor Processing Unit (TPU).

Google designed the TPU as an application-specific integrated circuit for accelerating neural-network workloads, especially operations used by its TensorFlow machine-learning software. Google first publicly described the TPU at the 2016 Google I/O conference.

Unlike a general-purpose CPU, a TPU is specialized for tensor operations, including the large matrix calculations common in neural networks. Google has used successive TPU generations in data centers for services such as search, translation, speech recognition, and image processing.

A common mix-up is calling every AI accelerator a TPU. TPU is specifically Google’s branded processor family; GPUs and other neural-processing chips are separate technologies. Google later introduced TPU generations with greater performance and expanded support for training as well as inference.