Tensor Processing Unit

Google's own accelerator chip for training and serving machine learning models built on frameworks such as JAX and PyTorch.

In more detail

TPUs suit large matrix-heavy workloads and are offered through Compute Engine, GKE and Gemini Enterprise Agent Platform. A GPU is the more general choice when a framework or library has no TPU support. Availability varies by zone, so check it before you plan capacity.

Related terms

Taught in

References

Certification courses