AI accelerators (Google TPU and successors)

Custom chips for neural-network math; Google's first TPU ran 15–30x faster than contemporary CPUs and GPUs, spawning many rivals.

Chips & processorsHardware
  1. 2016

    First version

    Google

    Image: Google · Unclear; Google's image page provides n… · Google Cloud Blog
  2. 2017
    TPU v2

    TPU v2

    Adds floating-point training, HBM and chip-to-chip networking for pods.

    Image: Rob Pegoraro · CC BY-NC-SA 2.0 · Flickr (via Openverse)
  3. 2021
    TPU v4

    TPU v4

    3D pod network with optical circuit switching; SparseCores added.

    Image: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan, Lifeng Na… · CC BY 4.0 · Wikimedia Commons

Contested: Google says ASIC work became urgent in 2013; TPUs ran in its datacenters before the 2016 announcement (often dated 2015).

See it on the timeline