Hardware · Established · Advanced

TPU (2)

Also known as: Tensor Processing Unit

Google's custom accelerators designed specifically for neural network training and inference.

What TPU (2) is

TPUs are application-specific chips optimised for large matrix operations, deployed in pods with high-speed interconnects for very large training runs.

How it works

Workloads reach them through JAX or TensorFlow via the XLA compiler, and they are available to external users through Google Cloud.

Why it matters

They are one of the few production alternatives to GPU-based training at frontier scale.

Common uses

→Large model training
→High-volume inference on Google Cloud

Strengths

✓Strong performance per watt on matrix workloads
✓Tight pod interconnect

Watch for

✓Available only through Google Cloud
✓Narrower framework support

Continue exploring

GPU for AI AI Compute JAX (2)Google