Infrastructure
aka Tensor Processing Unit
TPU
Google's custom ASIC for neural-network training and inference, available on GCP.
Definition
TPUs are Google's domain-specific accelerators optimised for dense matmul workloads. TPU v5p and Trillium (v6) train frontier models like Gemini. They are programmed primarily via JAX or TensorFlow and are exclusive to Google Cloud.
Common use cases
- Frontier training
- GCP inference
- Research on Google Cloud