Ironwood is Google's seventh-generation TPU, listed in Google's documentation as TPU7x. Each chip has 192 GiB of HBM, and a pod scales to 9,216 chips.
Specs from Google
Google's TPU7x page gives:
| Item | Value |
|---|---|
| HBM per chip | 192 GiB |
| HBM bandwidth per chip | 7380 GBps (about 7.37 TB/s) |
| Peak FP8 per chip | 4614 TFLOPs |
| Peak BF16 per chip | 2307 TFLOPs |
| Chips per pod | 9,216 (144 cubes) |
The page notes these are physical figures, and that monitoring tools and frameworks report less allocatable HBM.
What Google says it is for
Google's announcement says Ironwood is "purpose-built" for large-scale training, reinforcement learning and high-volume, low-latency inference. It claims a 10X peak performance improvement over TPU v5p and more than 4X better performance per chip than TPU v6e for training and inference, and a superpod with 1.77 petabytes of shared HBM.
Who uses it is covered in whether Anthropic uses Google TPUs. The competitor overview is NVIDIA's main competitors in AI.
