Edited by 
Short answer
Ironwood is Google's seventh-generation Tensor Processing Unit (TPU v7), announced in April 2025 and described by Google as its first TPU designed for the age of inference. Each chip delivers about 4,614 TFLOPS of FP8 compute with 192 GB of HBM at 7.37 TB/s, and chips are joined into pods of up to 9,216 chips for about 42.5 exaflops. Google says it doubles performance per watt over the previous generation, Trillium. It became generally available on Google Cloud in late 2025, and Google followed it with eighth-generation TPUs in 2026.
What is a TPU?
A Tensor Processing Unit is Google's own AI accelerator. Instead of the general-purpose design of a GPU, a TPU is built around large matrix multiply units (MXUs) using systolic arrays, where data pulses through a grid of multiply-accumulate units. Google has used TPUs internally since 2015 for Search, Translate and its Gemini models, and rents them through Google Cloud. Our TPU vs GPU guide explains the design differences in detail.
Ironwood specifications
| Specification | Ironwood (TPU v7), per chip |
|---|---|
| Announced | April 2025, Google Cloud Next |
| Peak compute | about 4,614 TFLOPS (FP8) |
| Memory | 192 GB HBM |
| Memory bandwidth | 7.37 TB/s |
| Inter-chip interconnect (ICI) | 1.2 TB/s bidirectional |
| Pod sizes | 256 chips or 9,216 chips |
| Full-pod compute | about 42.5 exaflops (FP8) |
| Cooling | Liquid cooled |
Ironwood vs Trillium (TPU v6e)
Google compared Ironwood with Trillium, the sixth-generation TPU launched in 2024:
- 6x the memory: 192 GB of HBM per chip versus 32 GB.
- About 4.5x the memory bandwidth: 7.37 TB/s per chip.
- 2x performance per watt.
- Faster chip-to-chip links, so large models can be split across many chips with less waiting.
Ironwood is also the first TPU to support FP8 natively, which matters because FP8 roughly doubles throughput over BF16 for inference. See our guide to precision formats.
Why Ironwood was built for inference
Training a frontier model is a huge one-off job; serving it to millions of users is a continuous cost that, for popular models, grows larger than training. Reasoning models and AI agents make this worse because they generate many more tokens per request.
Inference, especially token generation, is limited by memory bandwidth and capacity. Ironwood's large HBM capacity lets bigger models and longer contexts fit on fewer chips, and its bandwidth speeds up token generation. Ironwood also includes an enhanced SparseCore, a specialised unit for the embedding lookups used in ranking and recommendation models, which are a major workload at Google.
Ironwood pods
A single TPU is less important than how many work together. Ironwood chips connect through Google's inter-chip interconnect in a 3D torus, and racks are linked with optical circuit switches that can reconfigure the network.
A full 9,216-chip pod has about 1.77 petabytes of shared HBM. Google's Pathways software lets one program span the whole pod, or many pods, as if it were one large computer.
Ironwood vs NVIDIA Blackwell
| Ironwood (per chip) | NVIDIA B200 (per GPU) | |
|---|---|---|
| Peak FP8 (dense) | about 4.6 PFLOPS | about 4.5 PFLOPS |
| Memory | 192 GB HBM | 192 GB HBM3E |
| Memory bandwidth | 7.37 TB/s | 8 TB/s |
| Scale-up domain | Up to 9,216 chips (ICI) | 72 GPUs per NVLink domain (GB200 NVL72) |
| Availability | Google Cloud only | Every major cloud and on-premises |
| Software | JAX, PyTorch/XLA, vLLM TPU support | CUDA and the full ecosystem |
Per chip, Ironwood and Blackwell are in the same class. The differences are in scale-up design, power efficiency claims, software and how you buy them: TPUs are only rented through Google Cloud, while NVIDIA GPUs are sold everywhere. For the bigger picture of NVIDIA's flagship systems, see GB200 NVL72 explained.
Who uses Ironwood?
Google uses TPUs to train and serve its Gemini models. In October 2025, Anthropic announced an expansion of its Google Cloud agreement covering access to up to one million TPUs, one of the largest AI compute deals announced at the time. Other Google Cloud customers rent Ironwood for training and serving their own models.
What came after Ironwood?
Google kept a fast cadence. In April 2026 it introduced its eighth-generation TPUs, split into a training-focused chip (TPU 8t) and an inference chip (TPU 8i). In November 2026 it added two variants: the confidential-computing TPU 8tx and the dual-die TPU 8i Max. Read our news report on TPU 8tx and TPU 8i Max. Ironwood remains an important part of Google Cloud's AI infrastructure.
Frequently asked questions
What does Ironwood mean?
It is Google's code name for TPU v7. Google names TPU generations, such as Trillium (v6e) and Ironwood (v7).
Can I buy an Ironwood TPU?
No. TPUs are only available to rent through Google Cloud.
Is Ironwood only for inference?
No. Google designed it with inference as the main focus, but it can also train large models.
Sources and further reading
- Google, Ironwood: The first Google TPU for the age of inference (April 2025)
- Google Cloud, TPU documentation
- Anthropic, Expanding our use of Google Cloud TPUs and services (October 2025)
- TechCrunch, Google broadens its AI hardware portfolio with two new specialized chips (November 2026)
Leave a Reply