Why it matters

TPU is Google's AI hardware. Understanding shapes GCP AI strategy.

Advertisement

The architecture

MXU (matrix mult unit): systolic array.

HBM.

Fast inter-chip via ICI (torus).

TPU architectureMXU systolic arraymatrix mult accelHBMon-chip memICI + Toruschip-to-chipJAX + Flax favored; PyTorch via XLA works
TPU stack.
Advertisement

How it works end to end

Pods: many TPU chips + fast interconnect.

JAX: preferred framework. PyTorch via XLA.

Availability: GCP only.

Cost: competitive with GPU for large training.