TPU7x (Ironwood) performance optimizations

Fields

id

sources/alt-tpu-ironwood-perf

URL

https://cloud.google.com/tpu/docs/ironwood-performance

Publisher

Google Cloud

Kind

spec-sheet

Published

2026-10-02

Accessed

2026-10-04

notes

FP8 training (1.3x over BF16), sharding strategy framework (FSDP/TP/EP/CP), SparseCore collective offloading, activation rematerialization (25-30% overhead), scoped VMEM tuning (64 MiB per TensorCore), Tokamax kernels (Splash attention, Megablox GMM).

Notes

FP8 training (1.3x over BF16), sharding strategy framework (FSDP/TP/EP/CP), SparseCore collective offloading, activation rematerialization (25-30% overhead), scoped VMEM tuning (64 MiB per TensorCore), Tokamax kernels (Splash attention, Megablox GMM).

Cross-references

This record references no other records.