Definition
A tensor processing unit (TPU) is an application-specific integrated circuit (ASIC) designed by Google for machine learning workloads, centered on accelerating the matrix multiplications that dominate neural networks. It is offered primarily as a service through Google Cloud.
How it works
A TPU computes tensor operations on systolic arrays of matrix-multiply units, using lower-precision formats such as bfloat16 to raise throughput and energy efficiency. It gives up generality to make one class of computation extremely efficient.
Usage
TPUs are used through Google Cloud to train and serve large models, recommendation systems, and other large-scale machine learning workloads. They suit teams already building in the Google ecosystem, for example with JAX or TensorFlow.
Related and confusable terms
A GPU is a broader parallel processor, while a TPU is a tensor-specific ASIC with a narrower scope; a CPU is the general-purpose processor for serial logic. All three often divide the work within a single AI infrastructure stack.