If you're interested in Gemini Enterprise Agent Platform training clusters, contact your sales representative for access.

Gemini Enterprise Agent Platform training clusters supports a variety of
machine types to accommodate different workloads. You can choose from the
following options when configuring your cluster node pools.

## Accelerator-optimized machine types

| Machine type | GPU | GPUs per node |
|---|---|---|
| `a4x-highgpu-4g` | NVIDIA GB200 | 4 |
| `a4-highgpu-8g` | NVIDIA B200 | 8 |
| `a3-ultragpu-8g` | NVIDIA H200 | 8 |
| `a3-megagpu-8g` | NVIDIA H100 | 8 |
| `a3-highgpu-8g` | NVIDIA H100 | 8 |
| `g4-standard-48` | NVIDIA RTX PRO 6000 Blackwell | 1 |
| `g4-standard-96` | NVIDIA RTX PRO 6000 Blackwell | 2 |
| `g4-standard-192` | NVIDIA RTX PRO 6000 Blackwell | 4 |
| `g4-standard-384` | NVIDIA RTX PRO 6000 Blackwell | 8 |

> [!NOTE]
> **Note:** The `a3-highgpu-8g` machine type requires a separate allowlist, in addition to the one that grants access to Gemini Enterprise Agent Platform training clusters. To request access to A3 High, contact your sales representative.

## CPU and HPC machine types

| Machine type | Description |
|---|---|
| N2 family: `n2-standard-*`, `n2-highmem-*`, `n2-highcpu-*` | General-purpose nodes for data preprocessing, job orchestration, and other work that doesn't require an accelerator. N2 is also supported for Slurm login nodes. |
| N4 family: `n4-standard-*`, `n4-highmem-*`, `n4-highcpu-*` | General-purpose nodes for the same workloads as N2. You can also use N4 for Slurm login nodes. N4 supports only Hyperdisk Balanced (`hyperdisk-balanced`) boot disks. If you don't set `boot_disk.boot_disk_type`, `hyperdisk-balanced` is used. |
| N4D family: `n4d-standard-*`, `n4d-highmem-*`, `n4d-highcpu-*` | General-purpose AMD EPYC nodes for the same workloads as N2. N4D is also supported for Slurm login nodes. N4D supports only Google Cloud Hyperdisk Balanced (`hyperdisk-balanced`) boot disks, which is the default when `boot_disk.boot_disk_type` isn't set. |
| C4 family: `c4-standard-*`, `c4-highmem-*`, `c4-highcpu-*` | General-purpose nodes for compute-intensive preprocessing, orchestration, and CPU workloads. C4 is also supported for Slurm login nodes. C4 supports only Google Cloud Hyperdisk Balanced (`hyperdisk-balanced`) boot disks, which is the default when `boot_disk.boot_disk_type` isn't set. The `-lssd` and `-metal` variants aren't supported. |
| C4D family: `c4d-standard-*`, `c4d-highmem-*`, `c4d-highcpu-*` | General-purpose AMD EPYC nodes for compute-intensive preprocessing, orchestration, and CPU workloads. C4D supports only Google Cloud Hyperdisk Balanced (`hyperdisk-balanced`) boot disks, which is the default when `boot_disk.boot_disk_type` isn't set. The `-lssd` and `-metal` variants aren't supported. |
| C4A family: `c4a-standard-*`, `c4a-highmem-*`, `c4a-highcpu-*` | Arm-based (Axion) nodes, supported for Slurm login nodes only, and only in clusters whose compute node pools are also Arm-based. The `-lssd` and `-metal` variants aren't supported. |
| H4D family: `h4d-standard-*`, `h4d-highmem-*` | HPC-optimized nodes with RDMA networking, for tightly coupled simulation and preprocessing workloads. |

### C4A login nodes

C4A login nodes are Arm-based and are supported only in clusters where every
compute node pool is also Arm-based, such as
[A4X](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/training-clusters/compute-resources.md.txt#a4x-machine-type). A cluster that pairs a C4A login node with an x86
compute node pool, such as N2 or A3, is rejected when you create it.

Because the login node and the compute nodes share the same architecture,
binaries and container images you build on the login node run on the compute
nodes you submit them to.

C4A only supports a `hyperdisk-balanced` boot disk, which is the default.

## A4X machine type

Gemini Enterprise Agent Platform training clusters support the
[A4X accelerator-optimized machine type](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-machine-type)
(a4x-highgpu-4g), an exascale platform based on the
[NVIDIA GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/)
rack-scale architecture.

### Architectural comparison

The following table outlines the fundamental hardware differences between the
A4X family and other accelerator-optimized families.

| Feature | A4X (a4x-highgpu-4g) | A3 / A4H |
|---|---|---|
| **CPU Architecture** | ARM | X86 |
| **GPU Count** | 4 GPUs per node | 8 GPUs per node |
| **Reservation Type** | All capacity mode | Managed Mode |
| **Placement Policy** | Strict (Compact) | Flexible |

### A4X specific guidelines

- Your A4X node pool VM count must be a multiple of 18 (for example, 18, 36, 54). This is required because A4X capacity is provisioned in fixed, non-shareable 18-node blocks called NVLink domains. These domains are bound by a strict Compact Placement Policy, and any partially allocated blocks cannot be used by other clusters.
- Due to the ARM-based architecture of A4X nodes, you must make two key changes to your training workloads:
  - Use ARM-Compatible images: All training jobs must use a container image that is built for the ARM architecture.
  - Adapt for 4 GPUs: Your distributed training logic must be updated to correctly recognize and use the 4 GPUs available on each A4X node.
- Host Fault Reporting process and downtime When you report a host as faulty, be aware of the following recovery process:
  - No standby capacity: The system doesn't use a standby spare pool for an instant node replacement.
  - Repair-based recovery: The node remains unavailable until the underlying physical host is repaired.
  - Extended downtime: This repair process typically takes 3 to 14 days.

## Capacity provisioning

Choosing the right provisioning model is critical for balancing cost, speed,
and resource availability. See the following provisioning options:

- `RESERVATION`: Allocates nodes from a specific Compute Engine reservation
  that you've created in advance. This model ensures capacity and is the
  recommended choice for high-demand resources.

- `FLEX_START`: Utilizes the Dynamic Workload Scheduler to queue your job. The job begins
  automatically as soon as the requested compute resources become available,
  offering a flexible start time without requiring a reservation.

- `SPOT`: Provisions the node pool using Spot VMs. This is the most
  cost-effective option, but it should only be used for workloads that are
  fault-tolerant and can handle interruptions, as the VMs may be preempted at
  any time.

- `ON_DEMAND`: This is the default option for CPU-only node pools and is best
  suited for machine types that are not scarce. It provides standard VM
  instances with predictable, pay-as-you-go pricing.

Use the following guidance to make your selection:

- For high-demand GPU resources (like A3 and A4): The `RESERVATION` model is
  strongly recommended. It ensures you have dedicated access to the capacity
  you need for critical training jobs.

- For bursty or flexible workloads: Consider `FLEX_START` or `SPOT`.
  `FLEX_START` queues your job until resources are available, while `SPOT`
  offers significant cost savings for fault-tolerant jobs that can handle
  preemption.

- For abundant machine types: The `ON_DEMAND` model is the preferred choice.
  Use it for machine types that are not scarce and where immediate availability
  isn't a concern.

## Using a shared reservation (optional)

If you'd like to use a shared reservation, rather than a local reservation,
there are additional steps to take before you can create a cluster.

Before using a shared reservation with Gemini Enterprise Agent Platform training clusters, make sure
the shared
reservation works by manually creating a VM that uses the shared reservation.
If this VM creation works, move on to the next step.
In the cluster creation configuration, use the reservation name in the following
format:
`projects/RESERVATION_HOST_PROJECT_ID/zones/RESERVATION_ZONE/reservations/RESERVATION_NAME`.

## What's next

After selecting the compute and provisioning options for your training cluster,
you are ready to create the cluster and run a workload on it.

- Create a Compute Engine reservation: The `RESERVATION` model is used for allocating high-demand resources like GPUs. Learn how to create a new reservation in Compute Engine to get dedicated access to your required resources.
  - [Learn how to create a reservation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/use-reservations.md.txt)
- Create your training cluster: Apply the configurations you've learned about by following the step-by-step guide to create your first persistent training cluster using the Agent Platform API or `gcloud`.
  - [Learn how to create a training cluster](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/training-clusters/create-cluster.md.txt)
- Submit a training job to your cluster: Once your cluster is active, the next step is to run a workload. Submit a `CustomJob` that targets your persistent cluster for execution.
  - [Learn how to run a job on a training cluster](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/create-custom-job.md.txt)
- Adapt your code for distributed training: To take full advantage of a multi-node cluster, adapt your training code for a distributed environment.
  - [Learn about distributed training on Gemini Enterprise Agent Platform](https://docs.cloud.google.com/gemini-enterprise-agent-platform/machine-learning/training/distributed-training.md.txt)