This document provides an overview of Ironwood (TPU7x) in
Google Kubernetes Engine (GKE). Ironwood (TPU7x) is Google's seventh-generation
[Tensor Processing Unit (TPU)](https://docs.cloud.google.com/tpu), custom-designed for large-scale AI
workloads. It offers a significant performance improvement over previous TPU
generations, which lets you train and serve larger and more complex models.

## Characteristics of Ironwood (TPU7x)

Ironwood (TPU7x) introduces unique features that differentiate it from other TPU
versions. These features impact availability, node pool configuration, and
workload performance.

For information about the underlying hardware, see [Ironwood (TPU7x)
architecture](https://docs.cloud.google.com/tpu/docs/tpu7x).

### Availability

Ironwood (TPU7x) is available in GKE Standard clusters
that run version 1.34.0-gke.2201000 and later, and in Autopilot
clusters that run version 1.34.1-gke.3084001 and later.

### Workload policy for multi-host node pools

Ironwood (TPU7x) uses a [workload
policy](https://docs.cloud.google.com/ai-hypercomputer/docs/placement-policy-and-workload-policy) to
configure the physical placement of the underlying infrastructure when you
create multi-host TPU slice node pools. You create a workload policy and then
apply it by using the `--placement-policy` flag. This policy replaces the
`--tpu-topology` flag used by other TPU versions.

A workload policy is a type of resource policy that lets you configure the
physical placement of infrastructure. Ironwood (TPU7x) supports the *High
throughput* workload policy. This policy colocates the TPU VMs to reduce network
latency and lets you define the maintenance strategy to minimize workload
disruptions.

### NUMA binding

The Ironwood (TPU7x) architecture includes the following elements:

- Each Ironwood (TPU7x) virtual machine (VM) contains four chips and two NICs.
- Each VM contains two [Non-Uniform Memory Access (NUMA)](https://en.wikipedia.org/wiki/Non-uniform_memory_access) nodes.
- The CPU, memory, and NICs resources are equally split between the two NUMA nodes.

Accessing resources across different NUMA nodes (cross-NUMA access) can
introduce performance bottlenecks in your workloads. Therefore, to optimize your
workload performance, GKE lets you deploy your workloads in a
[*multi-container* setup](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/tpus.md.txt#deploy-multi-container). This binds each container to the CPU, memory, and TPU
resources within a given NUMA node.

## Reference implementations of LLMs

To learn how to deploy large language models (LLMs) on Ironwood (TPU7x), see the
following reference implementations. You can use one of the following options
for cluster creation:

- **GKE XPK** : use Accelerated Processing Kit (XPK) to quickly create GKE clusters and run workloads for proofs-of-concept and testing. For more information, see the [XPK
  documentation](https://github.com/AI-Hypercomputer/xpk/blob/main/README.md).
- **GKE on Google Cloud CLI:** use the Google Cloud CLI to create your GKE cluster instance manually for precise customization or expansion of existing production GKE environments.

| LLM | GKE XPK | GKE on Google Cloud CLI |
|---|---|---|
| Llama 70b with BF16 and a `4x4x4` topology | [Pretrain llama3.1-70b workload on Ironwood GKE clusters with XPK](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/llama3.1-70b/8k-bf16-tpu7x-4x4x4/xpk) | [Pretrain llama3.1-70b workload on Ironwood GKE clusters with Kubernetes JobSet](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/llama3.1-70b/8k-bf16-tpu7x-4x4x4/k8s) |
| DeepSeek with BF16 and a `4x4x8` topology | [Pretrain deepseek3-671b workload on Ironwood GKE clusters with XPK](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/deepseek3-671b/4k-bf16-tpu7x-4x4x8/xpk) | [Pretrain deepseek3-671b workload on Ironwood GKE clusters with Kubernetes JobSet](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/deepseek3-671b/4k-bf16-tpu7x-4x4x8/k8s) |
| GPT-oss-120b with BF16 and a `4x4x4` topology | [Pretrain gpt-oss-120b workload on Ironwood GKE clusters with XPK](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/gpt-oss-120b/8k-bf16-tpu7x-4x4x4/xpk) | [Pretrain gpt-oss-120b workload on Ironwood GKE clusters with Kubernetes JobSet](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/gpt-oss-120b/8k-bf16-tpu7x-4x4x4/k8s) |
| Qwen3-235b-a22b with BF16 and a `4x8x8` topology | [Pretrain qwen3-235b-a22b workload on Ironwood GKE clusters with XPK](https://github.com/AI-Hypercomputer/tpu-recipes/tree/main/training/ironwood/qwen3-235b-a22b/4k-bf16-tpu7x-4x8x8) | Not available |

## What's next

- Learn how to [plan TPUs in GKE](https://docs.cloud.google.com/kubernetes-engine/docs/concepts/plan-tpus.md.txt).
- Learn how to [deploy TPUs in GKE](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/tpus.md.txt).
- Try the end-to-end tutorials for Ironwood (TPU7x):
  - [Run training workload with Ironwood (TPU7x) and reservations using
    Cloud Storage](https://github.com/AI-Hypercomputer/xpk/blob/main/docs/usage/tpu7x/recipes/reservation_gcs_bucket_recipe.md)
  - [Run training workload with Ironwood (TPU7x) with flex-start
    using Filestore
    storage](https://github.com/AI-Hypercomputer/xpk/blob/main/docs/usage/tpu7x/recipes/flex_filestore_recipe.md)
  - [Run training workload with Ironwood (TPU7x) and
    flex-start using Lustre
    storage](https://github.com/AI-Hypercomputer/xpk/blob/main/docs/usage/tpu7x/recipes/flex_lustre_recipe.md)