Supported compute resources for Cluster Director

This document provides a comprehensive overview of the compute resources that are available for Cluster Director clusters, including CPU-based machines, GPU-accelerated machines, and the pre-configured OS images. You use these resources to deploy and manage the underlying infrastructure for high performance computing (HPC), artificial intelligence (AI), and machine learning (ML) workloads.

CPU-based machines

Cluster Director supports general-purpose machine series, which are primarily CPU-based and suitable for HPC workloads and cluster management roles. These machines are ideal for cluster components that don't require GPU acceleration, such as login nodes, or for running CPU-bound computational tasks.

N2 machine series

Cluster Director supports the N2 machine series. This series offers a balance of price and performance suitable for a variety of workloads, such as CPU-bound computational tasks or general-purpose HPC workloads.

The N2 machine types are differentiated by the amount of memory configured per vCPU. These memory configurations are as follows:

• standard: 4 GB of system memory per vCPU

• highmem: 8 GB of system memory per vCPU

• highcpu: 1 GB of system memory per vCPU

Each family offers a range of machine sizes, from 2 to 128 vCPUs. For login nodes, you can only use N2 standard machine types with 32 or fewer vCPUs. Otherwise, you can use any N2 machine type. The following table summarizes the available N2 machine types in Cluster Director:

N2 standard

Machine types	vCPUs^*	Memory (GB)	Default egress bandwidth (Gbps)^‡	Tier_1 egress bandwidth (Gbps)^#
n2-standard-2	2	8	Up to 10	N/A
n2-standard-4	4	16	Up to 10	N/A
n2-standard-8	8	32	Up to 16	N/A
n2-standard-16	16	64	Up to 32	N/A
n2-standard-32	32	128	Up to 32	Up to 50
n2-standard-48	48	192	Up to 32	Up to 50
n2-standard-64	64	256	Up to 32	Up to 75
n2-standard-80	80	320	Up to 32	Up to 100
n2-standard-96	96	384	Up to 32	Up to 100
n2-standard-128	128	512	Up to 32	Up to 100

^* A vCPU is implemented as a single hardware thread, or logical core, on one of the available CPU platforms.
^‡ Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information, see Network bandwidth.
^# Supports high-bandwidth networking for larger machine types. For Windows OS images, the maximum network bandwidth is limited to 50 Gbps.

N2 high-mem

Machine types	vCPUs^*	Memory (GB)	Default egress bandwidth (Gbps)^‡	Tier_1 egress bandwidth (Gbps)^#
n2-highmem-2	2	16	Up to 10	N/A
n2-highmem-4	4	32	Up to 10	N/A
n2-highmem-8	8	64	Up to 16	N/A
n2-highmem-16	16	128	Up to 32	N/A
n2-highmem-32	32	256	Up to 32	Up to 50
n2-highmem-48	48	384	Up to 32	Up to 50
n2-highmem-64	64	512	Up to 32	Up to 75
n2-highmem-80	80	640	Up to 32	Up to 100
n2-highmem-96	96	768	Up to 32	Up to 100
n2-highmem-128	128	864	Up to 32	Up to 100

^* A vCPU is implemented as a single hardware thread, or logical core, on one of the available CPU platforms.
^‡ Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information, see Network bandwidth.
^# Supports high-bandwidth networking for larger machine types. For Windows OS images, the maximum network bandwidth is limited to 50 Gbps.

N2 high-cpu

Machine types	vCPUs^*	Memory (GB)	Default egress bandwidth (Gbps)^‡	Tier_1 egress bandwidth (Gbps)^#
n2-highcpu-2	2	2	Up to 10	N/A
n2-highcpu-4	4	4	Up to 10	N/A
n2-highcpu-8	8	8	Up to 16	N/A
n2-highcpu-16	16	16	Up to 32	N/A
n2-highcpu-32	32	32	Up to 32	Up to 50
n2-highcpu-48	48	48	Up to 32	Up to 50
n2-highcpu-64	64	64	Up to 32	Up to 75
n2-highcpu-80	80	80	Up to 32	Up to 100
n2-highcpu-96	96	96	Up to 32	Up to 100

^* A vCPU is implemented as a single hardware thread, or logical core, on one of the available CPU platforms.
^‡ Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information, see Network bandwidth.
^# Supports high-bandwidth networking for larger machine types. For Windows OS images, the maximum network bandwidth is limited to 50 Gbps.

GPU-accelerated machines

Cluster Director supports a specific set of accelerator-optimized machine series designed for large-scale AI, ML, and HPC workloads.

GPU machine types are available only in specific regions and zones. To review availability, see GPU locations.

A4X machine series

A4X accelerator-optimized machine types use NVIDIA GB200 Grace Blackwell Superchips (nvidia-gb200) and are ideal for foundation model training and serving.

A4X is an exascale platform based on NVIDIA GB200 NVL72. Each machine has two sockets with NVIDIA Grace CPUs with Arm Neoverse V2 cores. These CPUs are connected to four NVIDIA B200 Blackwell GPUs with fast chip-to-chip (NVLink-C2C) communication.

Tip: To use A4X machine types in your cluster, you must reserve capacity through your account team to create VMs. For more information about the ways you can create A4X VMs in Cluster Director clusters, see Choose a consumption option.

						Attached NVIDIA GB200 Grace Blackwell Superchips
Machine type	vCPU count^1	Instance memory (GB)	Attached Local SSD (GiB)	Physical NIC count	Maximum network bandwidth (Gbps)^2	GPU count	GPU memory^3 (GB HBM3e)
a4x-highgpu-4g	140	884	12,000	6	2,000	4	744

¹A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
²Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information about network bandwidth, see Network bandwidth.
³GPU memory is the memory on a GPU device that can be used for temporary storage of data. It is separate from the instance's memory and is specifically designed to handle the higher bandwidth demands of your graphics-intensive workloads.

A4 machine series

A4 accelerator-optimized machine types have NVIDIA B200 Blackwell GPUs (nvidia-b200) attached and are ideal for foundation model training and serving.

Tip: To use A4 machine types in your cluster, you must reserve capacity to create VMs, create Spot VMs, or create Flex-start VMs. For more information about the ways you can create A4 VMs in Cluster Director clusters, see Choose a consumption option.

						Attached NVIDIA B200 Blackwell GPUs
Machine type	vCPU count^1	Instance memory (GB)	Attached Local SSD (GiB)	Physical NIC count	Maximum network bandwidth (Gbps)^2	GPU count	GPU memory^3 (GB HBM3e)
a4-highgpu-8g	224	3,968	12,000	10	3,600	8	1,440

¹A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
²Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information about network bandwidth, see Network bandwidth.
³GPU memory is the memory on a GPU device that can be used for temporary storage of data. It is separate from the instance's memory and is specifically designed to handle the higher bandwidth demands of your graphics-intensive workloads.

A3 machine series

The A3 machine series is powered by NVIDIA H100 or H200 SXM GPUs and is suitable for a wide range of large model training and inference workloads.

Cluster Director supports the A3 machine types that are described in the following sections.

A3 Ultra machine type

A3 Ultra machine types have NVIDIA H200 SXM GPUs (nvidia-h200-141gb) attached and provides the highest network performance in the A3 series. A3 Ultra machine types are ideal for foundation model training and serving.

Tip: To use A3 Ultra machine types in your cluster, you must reserve capacity to create VMs, create Spot VMs, or create Flex-start VMs. For more information about the ways to create A3 Ultra VMs in Cluster Director clusters, see Choose a consumption option.

						Attached NVIDIA H200 GPUs
Machine type	vCPU count^1	Instance memory (GB)	Attached Local SSD (GiB)	Physical NIC count	Maximum network bandwidth (Gbps)^2	GPU count	GPU memory^3 (GB HBM3e)
a3-ultragpu-8g	224	2,952	12,000	10	3,600	8	1128

¹A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
²Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information about network bandwidth, see Network bandwidth.
³GPU memory is the memory on a GPU device that can be used for temporary storage of data. It is separate from the instance's memory and is specifically designed to handle the higher bandwidth demands of your graphics-intensive workloads.

A3 Mega machine type

A3 Mega machine types have NVIDIA H100 SXM GPUs and are ideal for large model training and multi-host inference.

Tip: To use A3 Mega machine types in your cluster, you must create Spot VMs or Flex-start VMs. For more information about the ways to create A3 Mega VMs in Cluster Director clusters, see Choose a consumption option.

						Attached NVIDIA H100 GPUs
Machine type	vCPU count^1	Instance memory (GB)	Attached Local SSD (GiB)	Physical NIC count	Maximum network bandwidth (Gbps)^2	GPU count	GPU memory^3 (GB HBM3)
a3-megagpu-8g	208	1,872	6,000	9	1,800	8	640

¹A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
²Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. For more information about network bandwidth, see Network bandwidth.
³GPU memory is the memory on a GPU device that can be used for temporary storage of data. It is separate from the instance's memory and is specifically designed to handle the higher bandwidth demands of your graphics-intensive workloads.

Operating system (OS) images

OS images are pre-configured, bootable disk images used to deploy virtual machine (VM) instances in Cluster Director. These images are Docker images that are pre-configured with ML frameworks and libraries to simplify the model creation and training for large-scale AI and ML workloads.

Slurm OS images for Cluster Director

When you deploy a Slurm cluster, Cluster Director uses a custom OS image that is automatically built by Cluster Toolkit. This process helps ensure that Cluster Director provisions every node in the cluster with a correct and consistent software stack, including the controller and compute nodes.

These custom OS images are extensions of the Ubuntu LTS Accelerator OS images and include all necessary system software for cluster and workload management.

Included software

The custom OS images for Cluster Director nodes include the following software components by default:

• Operating system: Ubuntu 22.04 LTS

• Orchestration: Slurm version 24.11.2 and its dependencies, such as MUNGE and MariaDB.

• Containerization tools: the NVIDIA enroot container runtime and NVIDIA pyxis, used for running containerized workloads on Slurm clusters.

• Drivers and libraries: NVIDIA driver version 570, CUDA Toolkit (version 12.8), and libraries for GPUDirect-RDMA, including ibverbs-utils and rdma-core.

• Parallel computing libraries: OpenMPI and PMIx for managing parallel processing tasks across the cluster.

• Google Cloud integrations: the Ops Agent for monitoring and logging, and Cloud Storage FUSE for accessing Cloud Storage buckets from the cluster nodes.

What's next?

• Supported networking services in Cluster Director

• Try out Cluster Director