AI Research Atlas

Ironwood (TPU v7)

Google · 9 April 2025

Ironwood is Google's seventh-generation TPU and the first designed mainly for inference, with 192 GB HBM per chip and pods of 9,216 chips that deliver 42.5 exaflops.

Built for large thinking models. Roughly 2x performance per watt versus Trillium (v6e), 6x the HBM capacity at 7.37 TB/s per chip, and a faster inter-chip interconnect. Google compares one pod against the El Capitan supercomputer.

Date
Wednesday, 9 April 2025
Lab
Google
Kind
infra
Access
closed API

Figures

MeasureValueMeasured by
Pod compute42.5 exaflops
9,216 chips; Google compares with El Capitan at 1.7 EF
company
HBM per chip192 GB
7.37 TB/s bandwidth
company

All figures are Google's own. Eighth-generation TPUs followed at Cloud Next 2026-04.

Sources

  1. blog.google/products/google-cloud/ironwood-tpu-age-of-inference/

This record was checked against its sources on 6 October 2026. How we check

Related