Core concept
SystolicArray + HBM. Massive per-chip perf. Interconnected via ICI + OCI into Pods.
Advertisement
How it works
Software: XLA + JAX or PyTorch/XLA. Not everything works out of box (CUDA-native has gaps).
Advertisement
Trade-offs + gotchas
Google's ML. Cost-sensitive inference (v5e). Massive training (v5p, v6 Pods).