TPUv7 Ironwood
Also known as Ironwood, TPU v7, TPUv7, v7x
In plain English
Ironwood is the seventh-generation TPU and the first one Google sells for other companies to run their own inference on.
Technical definition
TPUv7 Ironwood is the Google accelerator with two independent compute dies per chip, native FP8 hardware, a 256x256 MXU, roughly six times the HBM of Trillium, and a 3D torus ICI fabric.
Cost at 100 tok/s/user, FP8 8k1k
$0.181 per million tokens vs $0.222 (B200) and $0.276 (B300)
Engineering details
Ironwood drops the MegaCore convention of TPU v4 and v5p, where two cores shared one memory space. Its two dies run as separate logical devices joined by a die-to-die link, so frameworks see two devices per chip. Each chip carries two TensorCores and four third-generation SparseCores. The base pod unit is a 4x4x4 cube of 64 chips, and optical circuit switches stitch cubes into a 9,216-chip superpod. Ironwood has no native FP4, so the current comparison against NVIDIA uses FP8 on both sides.
Why it matters
This is the first generation where Google competes for outside inference workloads with chips that can be bought outright or rented. Anthropic alone committed to more than one million TPUs. The FP8-only compute path caps the comparison against Blackwell until TPUv8i adds native FP4.
How to read it in InferenceX
In the InferenceX Official Preview, Ironwood serving Qwen3.5 397B in FP8 costs about $0.181 per million tokens at 100 tokens per second per user, against $0.222 for B200 and $0.276 for B300. At 20 tokens per second per user it delivers 50.4% more tokens per dollar than B200 and 96.0% more than B300 using external TCO.