A custom-built AI chip from Google that was introduced in 2016 and used in Google data centers. Unlike the NVIDIA GPU, which was originally designed for graphics and later upgraded for general parallel processing, the Tensor Processing Unit (TPU) was designed entirely for the type of matrix multiplication AI neural networks require. Running software from Google's TensorFlow AI library, TPU circuits process low-precision math very quickly (see
TF32).
Within a year, NVIDIA added Tensor processing cores to its GPUs in addition to its native CUDA cores that were added for general-purpose parallel processing (see
Tensor core and
CUDA).
In 2025, Google introduced Ironwood, which was enhanced for inference, although still used for training. Liquid cooled, up to 9,216 Ironwood TPUs can be linked with Google's InterChip Interconnect (ICI).
Major Changes
In 2026, Google split its 8th-generation TPU into training and inference chips (TPU8t and TPU8i). Also in 2026, a decade after the TPU's introduction, Google made them available on the open market. This was a watershed moment in the AI world because TPUs cost less than NVIDIA's GPUs. Within the six months after learning Meta was purchasing TPUs and the TPUs-for-sale-to-everyone announcement, NVIDIA lost a combined $400 billion in market valuation. It also caused NVIDIA to ramp up product cycles. See
Google Cloud Platform.
TPUs in a Smartphone
Starting in 2019, Pixel phones began using Edge TPUs, which were later integrated into the phone's primary chip (SoC). See
Tensor chip.