TPU (Tensor Processing Unit)
Definition pending
We've cataloged "TPU (Tensor Processing Unit)" but haven't written a full definition yet. Definitions are hand-curated rather than auto-generated, so it takes time to cover every term.
Want this one prioritized? Email us and we'll bump it.
Practical example
TPU (Tensor Processing Unit) is Google's custom AI accelerator. Available only through Google Cloud — you can't buy one. TPU v5p delivers 459 TFLOPS BF16 per chip in a 8,960-chip pod (4.1 exaFLOPS). TPUs excel at training at massive scale; for inference, they're competitive but less common than GPUs because of the cloud-only availability.
Workflow example
If you're on GCP: TPUs make sense for training large models (70B+) at scale. For inference, GPUs (A100/H100) on GCP are typically cheaper per token served. Use TPU if: (1) you're already in the GCP ecosystem, (2) you're training from scratch at >10^23 FLOP scale, (3) JAX is your framework (TPUs + JAX = great developer experience). For the local AI operator, TPUs are irrelevant — they're cloud-only silicon.