Core concept

TPU v5e: 197 TFLOPS BF16, 16GB HBM, cheaper. v5p: 459 TFLOPS BF16, 95GB HBM, ~2x v4.

Advertisement

How it works

Google Cloud only. XLA/JAX programming. TPU Pods for multi-thousand-chip training.

Advertisement

Trade-offs + gotchas

Google-scale training. Foundation models. Anthropic uses TPU + AWS Trainium.