Series

TPU

Google Cloud TPUs from first principles — the hardware and compiler that make them work, then what it takes to serve a model on one.

TPU FoundationsWhat a TPU actually is and how the whole stack fits together, from a single chip to Pod-scale distributed training.
2 parts · 53 min →
Multi-Tenant LLM ServingDeploying an LLM for many tenants on a single TPU machine — one HBM pool, one scheduler, and no horizontal escape hatch.
10 parts · 81 min →