This document describes the Compute Engine instances that have pre-attached
NVIDIA GPUs in the accelerator-optimized machine family. These instances are
designed specifically for artificial intelligence (AI), machine learning (ML),
high performance computing (HPC), and graphics-intensive applications.

The accelerator-optimized machine family includes the following machine series:
A5X, A4X Max, A4X, A4, A3, A2, G4, and G2. Each machine type within a series has a
specific model and number of NVIDIA GPUs attached. You can also attach [some GPU
models](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-general-purpose#overview) to N1
general-purpose machine types.

For information about accelerator-optimized instances with attached TPUs, see
[TPU machines in accelerator-optimized machine
family](https://docs.cloud.google.com/compute/docs/tpus/tpu-machines).

## GPU models and machine series

The following table summarizes the hardware specifications, machine series
mappings, networking bandwidth, and storage capabilities of GPU models available
on Compute Engine:

| GPU model | Machine series | Architecture | GPU memory and bandwidth | GPUs per compute instance | Max network bandwidth | Local SSD | Interconnect | NVIDIA RTX Virtual Workstation (vWS) support |
| NVIDIA Vera Rubin | [A5X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a5x-instances) | Rubin | 288 GB HBM4 (22 TBps) | 4 (NVL72) | 7,200 Gbps | 48,000 GiB | NVLink Full Mesh (3,600 GBps) | No |
| NVIDIA GB300 | [A4X Max](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) | Blackwell Ultra | 279 GB HBM3e (8 TBps) | 4 (NVL72) | 3,600 Gbps | 12,000 GiB | NVLink Full Mesh (1,800 GBps) | No |
| NVIDIA GB200 | [A4X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) | Blackwell | 186 GB HBM3e (8 TBps) | 4 (NVL72) | 2,000 Gbps | 12,000 GiB | NVLink Full Mesh (1,800 GBps) | No |
| NVIDIA B200 | [A4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4-vms) | Blackwell | 180 GB HBM3e (8 TBps) | 8 | 3,600 Gbps | 12,000 GiB | NVLink Full Mesh (1,800 GBps) | No |
| NVIDIA H200 | [A3 Ultra](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) | Hopper | 141 GB HBM3e (4.8 TBps) | 8 | 3,600 Gbps | 12,000 GiB | NVLink Full Mesh (900 GBps) | No |
| NVIDIA H100 | [A3 Mega, High, Edge](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) | Hopper | 80 GB HBM3 (3.35 TBps) | 1, 2, 4, 8 | Up to 1,000 Gbps | Up to 6,000 GiB | NVLink Full Mesh (900 GBps) | No |
| NVIDIA A100 (80GB) | [A2 Ultra](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a2-vms) | Ampere | 80 GB HBM2e (1.9 TBps) | 1, 2, 4, 8 | 100 Gbps | Up to 3,000 GiB | NVLink Full Mesh (600 GBps) | No |
| NVIDIA A100 (40GB) | [A2 Standard](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a2-vms) | Ampere | 40 GB HBM2 (1.6 TBps) | 1, 2, 4, 8, 16 | 100 Gbps | Up to 3,000 GiB | NVLink Full Mesh (600 GBps) | No |
| NVIDIA RTX PRO 6000 | [G4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-series) | Blackwell | 96 GB GDDR7 (1.597 TBps) | 1/8, 1/4, 1/2, 1, 2, 4, 8 | 200 Gbps | Up to 3,000 GiB | PCIe Gen 5 | Yes |
| NVIDIA L4 | [G2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-vms) | Ada Lovelace | 24 GB GDDR6 (300 GBps) | 1, 2, 4, 8 | 100 Gbps | Up to 3,000 GiB | PCIe Gen 4 | Yes |
| NVIDIA V100 | [N1+V100](https://docs.cloud.google.com/compute/docs/gpus.md.txt#v100-gpus) | Volta | 16 GB HBM2 (900 GBps) | 1, 2, 4, 8 | 100 Gbps | Supported | NVLink Ring (300 GBps) | No |
| NVIDIA T4 (approaching [end of support](https://docs.cloud.google.com/compute/docs/eol/t4-eos.md.txt)) | [N1+T4](https://docs.cloud.google.com/compute/docs/gpus.md.txt#t4-gpus) | Turing | 16 GB GDDR6 (320 GBps) | 1, 2, 4 | 100 Gbps | Supported | PCIe Gen 3 | Yes |
| NVIDIA P4 (approaching [end of support](https://docs.cloud.google.com/compute/docs/eol/p4-eos.md.txt)) | [N1+P4](https://docs.cloud.google.com/compute/docs/gpus.md.txt#p4-gpus) | Pascal | 8 GB GDDR5 (192 GBps) | 1, 2, 4 | 100 Gbps | Supported | PCIe Gen 3 | Yes |
|---|---|---|---|---|---|---|---|---|

<br />

## Recommended machine series by workload type

The following section provides the recommended machine series based on your GPU
workloads. For more information about how to choose a compute strategy for AI/ML
and HPC workloads, see [Choose a compute
strategy](https://docs.cloud.google.com/ai-hypercomputer/docs/choose-strategy) in the
AI Hypercomputer documentation.

### AI/ML training

- **Frontier foundation models and mixture-of-experts (MoE)** : pre-training
  large transformer models, multimodal architectures, and
  [mixture-of-experts (MoE)](https://en.wikipedia.org/wiki/Mixture_of_experts)
  networks requires high
  [Tensor Core](https://www.nvidia.com/en-us/data-center/tensor-cores/)
  throughput, large
  [High Bandwidth Memory (HBM)](https://en.wikipedia.org/wiki/High_Bandwidth_Memory)
  capacities (up to 288 GB per GPU), and high-speed
  [NVLink](https://www.nvidia.com/en-us/data-center/nvlink/) interconnect
  bandwidth (up to 3,600 GBps) to minimize multi-node synchronization
  latency. Recommended machine series:

  - [A5X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a5x-instances) (NVIDIA Vera Rubin)
  - [A4X Max](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) (NVIDIA GB300)
  - [A4X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) (NVIDIA GB200)
  - [A4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4-vms) (NVIDIA B200)
  - [A3 Ultra](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) (NVIDIA H200)
  - [A3 Mega](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) (NVIDIA H100)

  For more information, see [Recommendations for pre-training
  models](https://docs.cloud.google.com/ai-hypercomputer/docs/choose-strategy#pre-train) in the
  AI Hypercomputer documentation.
- **Fine-tuning and moderate model training** : adapting existing
  foundation models by using techniques like
  [parameter-efficient fine-tuning (PEFT)](https://en.wikipedia.org/wiki/Fine-tuning_(deep_learning))
  and [Low-Rank Adaptation (LoRA)](https://en.wikipedia.org/wiki/LoRA_(machine_learning))
  requires sufficient GPU memory to hold model weights and gradients
  without requiring exascale multi-node clustering. Recommended machine
  series:

  - [A3 High](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) (NVIDIA H100)
  - [A2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a2-vms) (NVIDIA A100)
  - [G4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-series) (NVIDIA RTX PRO 6000)

  For more information, see [Recommendations for fine-tuning
  models](https://docs.cloud.google.com/ai-hypercomputer/docs/choose-strategy#fine-tuning) in the
  AI Hypercomputer documentation.

### AI/ML inference

- **Large model inference (more than 70 billion parameters)** : serving
  [large language models (LLMs)](https://en.wikipedia.org/wiki/Large_language_model)
  benefits from high
  [High Bandwidth Memory (HBM)](https://en.wikipedia.org/wiki/High_Bandwidth_Memory)
  capacity (141--288 GB per GPU) to fit model weights across fewer
  GPUs, reducing inter-chip communication overhead. Recommended machine
  series:

  - [A5X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a5x-instances) (NVIDIA Vera Rubin)
  - [A4X Max](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) (NVIDIA GB300)
  - [A4X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) (NVIDIA GB200)
  - [A4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4-vms) (NVIDIA B200)
  - [A3 Ultra](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) (NVIDIA H200)
- **High-throughput and real-time serving**: low-latency interactive
  queries, streaming responses, and image generation use
  hardware-accelerated quantized numeric formats, such as 8-bit
  floating-point (FP8), 8-bit integer (INT8), and 4-bit integer (INT4),
  to maximize request throughput per dollar. Recommended machine series:

  - [G2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-vms) (NVIDIA L4)
  - [G4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-series) (NVIDIA RTX PRO 6000)
  - [A3 Edge and
    High](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines#a3-vms) (NVIDIA H100)

  For more information, see [Recommendations for serving
  inference](https://docs.cloud.google.com/ai-hypercomputer/docs/choose-strategy#inference) in the
  AI Hypercomputer documentation.

### Graphics and visual

- **3D CAD, digital twins, and virtual workstations** : interactive 3D
  modeling,
  [computer-aided design (CAD)](https://en.wikipedia.org/wiki/Computer-aided_design),
  architectural rendering, and
  [digital twins](https://en.wikipedia.org/wiki/Digital_twin) (such as
  [NVIDIA Omniverse](https://www.nvidia.com/en-us/omniverse/)) use
  dedicated
  [ray tracing (RT) cores](https://en.wikipedia.org/wiki/Ray_tracing_(graphics))
  and
  [NVIDIA RTX Virtual Workstations (vWS)](https://www.nvidia.com/en-us/design-visualization/virtual-workstation/)
  for remote visual workstations and fractional virtual GPUs (vGPUs).
  Recommended machine series:

  - [G4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-series) (NVIDIA RTX PRO 6000)
  - [G2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-vms) (NVIDIA L4)
- **Video processing and transcoding** : live streaming pipelines, video
  transcoding, and media processing use dedicated hardware
  [NVIDIA Encoder (NVENC)](https://en.wikipedia.org/wiki/Nvidia_NVENC) and
  [NVIDIA Decoder (NVDEC)](https://en.wikipedia.org/wiki/Nvidia_NVDEC)
  video engines supporting [AV1](https://en.wikipedia.org/wiki/AV1),
  [High Efficiency Video Coding (HEVC)](https://en.wikipedia.org/wiki/High_Efficiency_Video_Coding),
  and [H.264](https://en.wikipedia.org/wiki/Advanced_Video_Coding) codecs.
  Recommended machine series:

  - [G2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-vms) (NVIDIA L4)
  - [G4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-series) (NVIDIA RTX PRO 6000)

### HPC and simulation

- **Scientific simulation and modeling** : applications in
  [molecular dynamics](https://en.wikipedia.org/wiki/Molecular_dynamics)
  (such as [GROMACS](https://www.gromacs.org/) and
  [AMBER](https://ambermd.org/)),
  [computational fluid dynamics (CFD)](https://en.wikipedia.org/wiki/Computational_fluid_dynamics),
  quantum chemistry, astrophysics, and weather forecasting require high
  64-bit double-precision (FP64) floating-point performance and high memory
  bandwidth. Recommended machine series:

  - [A5X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a5x-instances) (NVIDIA Vera Rubin)
  - [A4X](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms) (NVIDIA GB200)
  - [A4](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4-vms) (NVIDIA B200)
  - [A3](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-vms) (NVIDIA H100)
  - [A2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a2-vms) (NVIDIA A100)

  For more information, see [Recommendations for
  HPC](https://docs.cloud.google.com/ai-hypercomputer/docs/choose-strategy#recommendations_for_hpc) in the
  AI Hypercomputer documentation.

## The A5X machine series

> [!CAUTION]
> **Caution:** The [Compute Engine Service Level Agreement (SLA)](https://cloud.google.com/compute/sla) doesn't apply to the A5X machine series.

The A5X machine series runs on an exascale platform based on
[NVIDIA's Vera Rubin NVL72 rack-scale architecture](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/)
and is optimized for large-scale ML training, inference for reasoning models,
Mixture of Experts (MoE) workloads, and HPC tasks. A5X is available only as
[bare metal instances](https://docs.cloud.google.com/compute/docs/instances/bare-metal-instances.md.txt),
which provide direct access to the host server's CPU and memory,
without Compute Engine's hypervisor in the middle.

All machine types in the A5X series have two sockets with 88-core NVIDIA Vera
CPUs. These CPUs connect to four NVIDIA Rubin GPUs with fast chip-to-chip
([NVLink-C2C](https://www.nvidia.com/en-us/data-center/nvlink-c2c/))
communication, creating a unified, coherent memory space across the CPUs and
GPUs.

### NVLink domain

The A5X machine series is built on NVIDIA's NVL72 rack-scale architecture,
which uses NVLink domains to enable large-scale, high-performance GPU computing.
An NVLink domain is a group of interconnected NVIDIA NVSwitch chips and the
GPUs that connect to them, forming a high-speed network fabric that enables
direct and fast communication between GPUs. For the A5X machine type, a single
NVL72 (NVLink) domain consists of up to 18 instances and 72 GPUs.

### A5X features

A5X bare metal instances provide the following features:

- **GPU acceleration with NVIDIA Vera Rubin superchips**: A5X instances have
  NVIDIA Vera Rubin superchips automatically attached. These
  superchips feature NVIDIA Rubin GPUs, offering 288 GB of HBM4 memory per
  GPU, or 1,152 GB per instance, and up to 20.7 TB of total GPU memory
  per NVL72 domain.

- **Enhanced networking with RoCE** : For A5X instances, Remote Direct Memory
  Access (RDMA) over Converged Ethernet (RoCE) increases network performance by
  combining eight NVIDIA ConnectX-9 (CX-9) SuperNICs, with two per GPU, and
  Google's high-bandwidth, low-latency
  [Virgo network](https://cloud.google.com/blog/products/networking/introducing-virgo-megascale-data-center-fabric),
  which features
  [eight-way rail-alignment](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#multi-layer-arch).
  This configuration delivers up to 6,400 Gbps of GPU-to-GPU bandwidth,
  which is optimal for demanding large-scale training, reasoning, and HPC
  workloads. The Virgo network provides a non-blocking scale-out fabric with a mesh
  topology that connects instances within and across buildings. It scales
  clusters up to 80,000 Vera Rubin GPUs within a single site and up to
  960,000 GPUs across multiple sites.

  For storage and management networking, each instance also has two
  400 Gbps Titanium NICs that provide up to 800 Gbps of
  combined bandwidth. These NICs use the Infrastructure Data Path Function
  (IDPF) network interface.
- **CPU platform** : The A5X machine type uses the NVIDIA Vera CPU platform,
  with two 88-core sockets providing 176 vCPUs across 176 physical CPU
  cores and 1,440 GB of instance memory. For more details about the
  platform, see [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).

- **NVLink scalability**: Multi-node NVLink 6 scales up to 72 GPUs in a single
  domain and provides bidirectional GPU NVLink bandwidth of 3,600 GBps per
  GPU.

- **Disk support**: A5X instances include Local SSD disks for fast scratch
  storage to feed data into GPUs and prevent I/O bottlenecks. Each A5X instance
  automatically includes 48,000 GiB of PCIe Gen6 NVMe Local SSD storage
  across four 12,000 GiB partitions.

  For durable storage, you can attach up to 512 TiB of
  Hyperdisk storage. For more information about disk types, see
  [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **Dense allocation and topology-aware scheduling**: The A5X machine
  type supports requesting blocks and subblocks of densely allocated capacity.
  Compute Engine allocates your host machines physically close to each
  other. It provisions them as blocks of resources and interconnects them with a
  high bandwidth and low latency Virgo network fabric to minimize network hops
  and optimize for low latency. Additionally, for A5X instances, you can view topology information at
  the host, subblock, block, and cluster levels to optimize job placement.

### Pricing and consumption options for A5X instances

A5X instances support the following consumption and discount options:

- [Future reservations for capacity blocks](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt)
- Resource-based committed use discounts (CUDs)

Google Cloud bills A5X instances for their attached GPUs, predefined
vCPUs, memory, and bundled Local SSD disks. Google Cloud bills
Hyperdisk and network usage separately from machine type pricing.
For more information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning)
and [Accelerator-optimized machine type family pricing](https://cloud.google.com/compute/vm-instance-pricing#accelerator-optimized).

### A5X machine type (NVIDIA Vera Rubin)


[A5X accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a5x-instances)

machine types use NVIDIA Vera Rubin superchips and
are ideal for foundation model training, reasoning, and serving. A5X machine types are available
as [bare metal instances](https://docs.cloud.google.com/compute/docs/instances/bare-metal-instances.md.txt).


A5X is an exascale platform based on
[NVIDIA Vera Rubin
NVL72](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/). Each machine has two sockets with NVIDIA Vera CPUs. These CPUs connect to four NVIDIA Rubin GPUs with fast
chip-to-chip ([NVLink-C2C](https://www.nvidia.com/en-us/data-center/nvlink-c2c/))
communication. For information about machine type limitations, see
[A5X limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a5x-limitations).

> [!NOTE]
> **Note:** To get started with A5X instances, contact your account team to [reserve capacity](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt) and get instructions to create instances and clusters.

| Machine type | vCPU count | Instance memory (GB) | Attached Local SSD (GiB) | Physical NIC count | Maximum network bandwidth (Gbps) | NVIDIA Rubin GPU count | GPU memory (GB HBM4) |
| `a5x-highgpu-4g-metal` | 176 | 1,440 | 48,000 | 10 | 7,200 | 4 | 1,152 |
|---|---|---|---|---|---|---|---|

<br />

#### A5X machine type notes


The following notes apply to the preceding table:

- **vCPU count** : for A5X bare metal instances, the vCPU count of 176 represents the 176 physical CPU cores across two 88-core sockets on the NVIDIA Vera [CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU memory is the memory on a GPU device that stores temporary data. It's separate from the instance's memory and handles the higher bandwidth demands of your graphics-intensive workloads.

<br />

### A5X limitations

The following limitations apply to A5X instances:

> [!CAUTION]
> **Caution:** The [Compute Engine Service Level Agreement (SLA)](https://cloud.google.com/compute/sla) doesn't apply to the A5X machine series.

- You can only request capacity that uses the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning).
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) or [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use this machine type.
- You can only create instances in certain [regions and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones.md.txt#view-using-table).
- You can't use Persistent Disk (regional or zonal). You can only use [Google Cloud Hyperdisk](https://docs.cloud.google.com/compute/docs/disks/hyperdisks.md.txt).
- This machine type is available only on the [NVIDIA Vera platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#arm_processors).
- You can't change the machine type of an A5X instance. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on this machine type. For a list of supported Linux operating systems, review the [supported operating systems for GPU instances](https://docs.cloud.google.com/compute/docs/images/os-details.md.txt#gpu-support).
- You can't update the machine firmware, such as the IDPF network interface, NVIDIA ConnectX-9 SuperNIC, or Local SSD firmware, on A5X instances.
- A5X instances don't support the following features:
  - [Sole-tenancy](https://docs.cloud.google.com/compute/docs/nodes/sole-tenant-nodes.md.txt)
  - [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
  - [Flex-start VMs](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- You can't attach Hyperdisk ML disks created before February 4, 2026, to A5X instances.

### Supported disk types for A5X instances


A5X instances can use the following block storage types:

- Hyperdisk Balanced (`hyperdisk-balanced`): the only disk type supported for the boot disk
- Hyperdisk Balanced High Availability (`hyperdisk-balanced-high-availability`)
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Local SSD: four 12,000 GiB Local SSD disks come pre-attached to each A5X instance

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD |
| `a5x-highgpu-4g-metal` | 32 | 32 | 32 | 32 | 32 | 8 | 4 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Google Cloud charges for Hyperdisk usage separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Hyperdisk pricing](https://cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />

#### Disk and capacity limits


You can attach a mixture of different Hyperdisk types to
an instance, but the maximum total disk capacity (in TiB) across all disk
types can't exceed 512 TiB for all Hyperdisks.

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance).

<br />

<br />

### Maintenance experience for A5X instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
A5X

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| A5X | 4 | Minimum of 90 days | Terminates with [Local SSD data persistence](https://docs.cloud.google.com/compute/docs/instances/host-maintenance-overview.md.txt#disk-persistence) | 90 days | Yes | No |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

For more information about host maintenance in AI Hypercomputer, see
[Understand host maintenance](https://docs.cloud.google.com/ai-hypercomputer/docs/manage/host-events.md.txt#review)
in the AI Hypercomputer documentation.

## The A4X Max and A4X machine series

> [!CAUTION]
> **Caution:** The [Compute Engine Service Level Agreement (SLA)](https://cloud.google.com/compute/sla) doesn't apply to the A4X Max and A4X machine series.

The A4X Max and A4X machine series runs on an exascale platform based on
[NVIDIA's rack-scale architecture](https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/)
and is optimized for compute and memory-intensive, network-bound ML training and
HPC workloads. A4X Max and A4X differ primarily in their GPU and networking
components. A4X Max is available only as
[bare metal instances](https://docs.cloud.google.com/compute/docs/instances/bare-metal-instances.md.txt),
which provide direct access to the host server's CPU and memory,
without Compute Engine's hypervisor in the middle.

All machine types in the A4X Max and A4X series have two sockets with NVIDIA
Grace™ CPUs with Arm® Neoverse™ V2 cores. These CPUs connect to four GPUs with
fast chip-to-chip
[NVLink-C2](https://www.nvidia.com/en-us/data-center/nvlink-c2c/) communication.

### NVLink domain

Both A4X Max and A4X machine series are built on NVIDIA's NVL72 rack-scale
architecture, which uses NVLink domains to enable large-scale,
high-performance GPU computing.
An NVLink Domain is a group of interconnected NVIDIA NVSwitch chips and the
GPUs that connect to them, forming a high-speed network fabric that allows for
direct and fast communication between GPUs. For A4X Max and A4X machine types,
a single NVL72 (NVLink) Domain is composed of 18 instances and 72 GPUs.

### A4X Max and A4X comparison

The following table compares the features that differ between the A4X Max and
A4X machine types:

| Feature | A4X Max (NVIDIA GB300) | A4X (NVIDIA GB200) |
| GPU acceleration | A4X Max instances have NVIDIA GB300 Ultra Superchips automatically attached. These Superchips feature [NVIDIA B300 GPUs](https://www.nvidia.com/en-us/data-center/b300/), offering up to 20 TB of total GPU memory per NVL72 domain, which provides roughly 279 GB per GPU. | A4X instances have NVIDIA GB200 Superchips automatically attached. These Superchips feature [NVIDIA B200 GPUs](https://www.nvidia.com/en-us/data-center/b200/) and offer 186 GB of memory per GPU. |
| Enhanced networking with RoCE | For A4X Max instances, RDMA over Converged Ethernet (RoCE) increases network performance by combining [NVIDIA ConnectX-8 (CX-8) SuperNICs](https://resources.nvidia.com/en-us-accelerated-networking-resource-library/connectx-datasheet-c) and Google's datacenter-wide network, which features [eight-way rail-alignment](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#multi-layer-arch). This configuration delivers up to 3,200 Gbps of bandwidth, which is optimal for demanding large-scale training and HPC tasks. For general-purpose networking, each instance also has up to 400 Gbps of bandwidth. | For A4X instances, RDMA over Converged Ethernet (RoCE) increases network performance by combining [NVIDIA ConnectX-7 (CX-7) NICs](https://resources.nvidia.com/en-us-accelerated-networking-resource-library/connectx-7-datasheet) and Google's datacenter-wide network, which features [four-way rail-alignment](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#multi-layer-arch). This architecture provides up to 1,600 Gbps of bandwidth, enabling high-throughput, low-latency communication for large-scale distributed workloads. For general-purpose networking, each instance also has up to 400 Gbps of bandwidth. |
| Performance | The NVIDIA GB300 Ultra Superchips provide 15 PetaFLOPS of dense FP4 performance. For large-scale FP4 inference, the GB300 Ultra Superchips are expected to deliver 20--40% higher performance over the GB200 Superchips. | The NVIDIA GB200 Superchips provide 10 PetaFLOPS of dense FP4 performance. |
| Bare metal and VM support | Bare metal instances only | VM instances only |
| OS support | A4X Max instances support a range of Linux OS images. However, because bare metal instances use the Infrastructure Data Path Function (IDPF) network driver, your OS image must support IDPF. To use an OS image that is available on Compute Engine, see [OS images that support IDPF](https://docs.cloud.google.com/compute/docs/images/os-details.md.txt#networking). | A4X instances support a range of Linux OS images. For a complete list of supported operating systems on Compute Engine, see [OS support for GPUs](https://docs.cloud.google.com/compute/docs/images/os-details.md.txt#gpu_support). |
|---|---|---|

Both A4X Max and A4X machine types share the following features:

- **CPU platform** : Both A4X Max and A4X machine types use the NVIDIA Grace CPU
  platform with Arm® Neoverse™ V2 cores. For more details about the platform,
  see [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).

- **NVLink scalability**: Multi-node NVLink scales up to 72 GPUs in a single
  domain and provides bidirectional GPU NVLink bandwidth of 1,800 GBps per
  GPU.

- **Disk support**: A4X Max and A4X instances support Local SSD for fast scratch
  disks, which is useful for feeding data into GPUs while preventing I/O
  bottlenecks. 12,000 GiB of Local SSD is automatically added to A4X Max
  and A4X instances.

  For durable storage, you can attach up to 512 TiB of
  Hyperdisk storage. For more information about disk types, see
  [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **Dense allocation and topology-aware scheduling support**: Both A4X Max and
  A4X machine types support requesting blocks of densely allocated capacity.
  Your host machines are allocated physically close to each other, provisioned
  as blocks of resources, and interconnected with a dynamic ML network fabric to
  minimize network hops and optimize for low latency. Additionally, for A4X Max
  and A4X instances, you can get topology information at the node and cluster
  level that can be used for job placement.

### Pricing and consumption options for A4X Max and A4X instances

A4X Max and A4X instances support the following consumption and discount
options:

- [Future reservations for capacity blocks](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt)
- Resource-based committed use discounts (CUDs)

Google Cloud bills A4X Max and A4X instances for their attached GPUs,
predefined vCPUs, memory, and bundled Local SSD disks. Hyperdisk
and network usage are billed separately from machine type pricing. For more
information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning),
[Accelerator-optimized machine type family pricing](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized),
and [GPU pricing](https://docs.cloud.google.com/compute/gpus-pricing).

### A4X Max machine type (NVIDIA GB300)


[A4X Max accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms)

machine types use NVIDIA GB300 Grace Blackwell Ultra Superchips (`nvidia-gb300`) and
are ideal for foundation model training and serving. A4X Max machine types are available
as [bare metal instances](https://docs.cloud.google.com/compute/docs/instances/bare-metal-instances.md.txt).


A4X Max is an exascale platform based on
[NVIDIA GB300
NVL72](https://www.nvidia.com/en-us/data-center/gb300-nvl72/). Each machine has two sockets with NVIDIA Grace CPUs with Arm
Neoverse V2 cores. These CPUs are connected to four NVIDIA B300 Blackwell GPUs with fast
chip-to-chip ([NVLink-C2C](https://www.nvidia.com/en-us/data-center/nvlink-c2c/))
communication. For information about machine type limitations, see
[A4X Max and A4X limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-limitations).

> [!NOTE]
> **Note:** When provisioning A4X Max instances, you must [reserve capacity](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt) to create instances and clusters. You can then create instances that use the features and services available from AI Hypercomputer. For more information, see [Deployment options overview](https://docs.cloud.google.com/ai-hypercomputer/docs/create/create-overview.md.txt) in the AI Hypercomputer documentation.

| Machine type | vCPU count | Instance memory (GB) | Attached Local SSD (GiB) | Physical NIC count | Maximum network bandwidth (Gbps) | NVIDIA GB300 GPU count | GPU memory (GB HBM3e) |
| `a4x-maxgpu-4g-metal` | 144 | 960 | 12,000 | 6 | 3,600 | 4 | 1,116 |
|---|---|---|---|---|---|---|---|

<br />

#### A4X Max machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

### A4X machine type (NVIDIA GB200)


[A4X accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-vms)

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](https://www.nvidia.com/en-us/data-center/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](https://www.nvidia.com/en-us/data-center/nvlink-c2c/))
communication. For information about machine type limitations, see
[A4X Max and A4X limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4x-limitations).

> [!NOTE]
> **Note:** When provisioning A4X instances, you must [reserve capacity](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt) to create instances and clusters. You can then create instances that use the features and services available from AI Hypercomputer. For more information, see [Deployment options overview](https://docs.cloud.google.com/ai-hypercomputer/docs/create/create-overview.md.txt) in the AI Hypercomputer documentation.

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

<br />

#### A4X machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

### A4X Max and A4X limitations

The following limitations apply to A4X Max and A4X instances:

### A4X Max instances


> [!CAUTION]
> **Caution:** The [Compute Engine Service Level Agreement (SLA)](https://cloud.google.com/compute/sla) doesn't apply to the A4X Max machine series.

- You can only request capacity that uses the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning).
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) or [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use this machine type.
- You can only create instances in certain [regions and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones.md.txt#view-using-table).
- You can't use Persistent Disk (regional or zonal). You can only use [Google Cloud Hyperdisk](https://docs.cloud.google.com/compute/docs/disks/hyperdisks.md.txt).
- This machine type is only available on the [NVIDIA Grace platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#arm_processors).
- Machine type changes aren't supported for A4X Max. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on this machine type. For a list of supported Linux operating systems, review the [supported operating systems for GPU instances](https://docs.cloud.google.com/compute/docs/images/os-details.md.txt#gpu-support).
- A4X Max instances don't support the following:
  - [Sole-tenancy](https://docs.cloud.google.com/compute/docs/nodes/sole-tenant-nodes.md.txt)
  - [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
  - [Flex-start VMs](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- You can't attach Hyperdisk ML disks created before February 4, 2026 to A4X Max machine types.

<br />

### A4X instances


> [!CAUTION]
> **Caution:** The [Compute Engine Service Level Agreement (SLA)](https://cloud.google.com/compute/sla) doesn't apply to the A4X machine series.

- You can only request capacity that uses the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning).
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) or [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use this machine type.
- You can only create instances in certain [regions and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones.md.txt#view-using-table).
- You can't use Persistent Disk (regional or zonal). You can only use [Google Cloud Hyperdisk](https://docs.cloud.google.com/compute/docs/disks/hyperdisks.md.txt).
- This machine type is only available on the [NVIDIA Grace platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#arm_processors).
- Machine type changes aren't supported for A4X. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on this machine type. For a list of supported Linux operating systems, review the [supported operating systems for GPU instances](https://docs.cloud.google.com/compute/docs/images/os-details.md.txt#gpu-support).
- For A4X instances, when you use `ethtool -S` to monitor GPU networking, the physical port counters that end in `_phy` don't update. This is expected behavior for instances that use the MRDMA Virtual Function (VF) architecture. For more information, see [MRDMA functions and network monitoring tools](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#mrdma-monitoring).
- A4X instances don't support the following:
  - [Sole-tenancy](https://docs.cloud.google.com/compute/docs/nodes/sole-tenant-nodes.md.txt)
  - [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
  - [Flex-start VMs](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- You can't attach Hyperdisk ML disks created before February 4, 2026 to A4X machine types.

<br />

### Supported disk types for A4X Max and A4X instances


### A4X Max

A4X Max instances can use the following block storage types:

- Hyperdisk Balanced (`hyperdisk-balanced`): this is the only disk type that is supported for the boot disk
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Local SSD: which is automatically added to instances that are created by using any of the A4X Max machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD |
| `a4x-maxgpu-4g-metal` | 32 | 32 | 32 | 32 | 8 | 4 |
|---|---|---|---|---|---|---|

### A4X

A4X instances can use the following block storage types:

- Hyperdisk Balanced (`hyperdisk-balanced`): this is the only disk type that is supported for the boot disk
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Hyperdisk ML (`hyperdisk-ml`)
- Local SSD: which is automatically added to instances that are created by using any of the A4X machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD |
| `a4x-highgpu-4g` | 128 | 128 | 0 | 0 | 128 | 8 | 4 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk usage is charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />

#### Disk and capacity limits


You can attach a mixture of different Hyperdisk types to
an instance, but the maximum total disk capacity (in TiB) across all disk
types can't exceed 512 TiB for all Hyperdisks.

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance).

<br />

<br />

### Maintenance experience for A4X Max and A4X instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
A4X Max and A4X

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| A4X Max and A4X | 4 | Minimum of 90 days | Stop and restart (with [Local SSD data persistence](https://docs.cloud.google.com/compute/docs/instances/host-maintenance-overview.md.txt#disk-persistence)) | 90 days | Yes | No |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

For more information about host maintenance in AI Hypercomputer, see
[Understand host maintenance](https://docs.cloud.google.com/ai-hypercomputer/docs/manage/host-events.md.txt#review)
in the AI Hypercomputer documentation.

## The A4 machine series

The A4 machine series offers machine types with up to 224 vCPUs, and
3,968 GB of memory. A4 instances provide up to 3x performance of previous
GPU instance types for most GPU accelerated workloads. A4 is recommended for ML
training workloads especially at large scales---for example,
hundreds or thousands of GPUs. The A4 machine series is available in a single
machine type.

VM instances created by using the A4 machine type provide the following
features:

- **GPU acceleration with NVIDIA B200 GPUs** :
  [NVIDIA B200 GPUs](https://www.nvidia.com/en-us/data-center/b200/)
  are automatically attached to A4 instances,
  which offer 180 GB GPU memory per GPU.

- **5th Generation Intel Xeon Scalable Processor (Emerald Rapids)** : offers up
  to 4.0 GHz sustained single-core max turbo frequency. For more information
  about this processor, see [CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).

- **NVLink scalability** : NVIDIA B200 GPUs provide
  [GPU NVLink bandwidth](https://www.nvidia.com/en-us/data-center/nvlink/)
  of 1,800 GBps, bidirectionally per GPU.

  With all-to-all NVLink topology between 8 GPUs in a system, the
  aggregate NVLink Bandwidth is up to 14.4 TBps.
- **Enhanced networking with RoCE** : RDMA over Converged Ethernet (RoCE) increases the network performance
  by combining [NVIDIA ConnectX-7 network interface cards (NICs)](https://www.nvidia.com/content/dam/en-zz/Solutions/networking/ethernet-adapters/connectx-7-datasheet-Final.pdf)
  with Google's datacenter-wide four-way rail-aligned network. By leveraging
  RDMA over Converged Ethernet (RoCE), A4 instances achieve much higher throughput between instances in a
  cluster compared to most A3 instances, except those running on the
  A3 Ultra machine type.

  > [!NOTE]
  > **Note:** Because of the difference in network topology, you can't move workloads between A4 instances and instances created by using any of the previous generation accelerator-optimized machine series. You can only move workloads between instances running on A4 and A3 Ultra machine types.

- **Increased network speeds**: Offers up to 4x networking speeds when compared
  to the previous generation A2 instances.

  For more information about networking, see
  [Network bandwidths and GPUs](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#a3-bandwidth).
- **Virtualization optimizations for data transfers and recovery**:
  the Peripheral Component Interconnect Express (PCIe) topology of A4 instances
  provides more accurate locality information that workloads can use to optimize
  data transfers.

  The GPUs also expose Function Level Reset (FLR) for graceful recovery
  from failures and atomic operations support for concurrency improvements
  in certain scenarios.
- **Disk support**: A4 instances support Local SSD for fast scratch disks,
  which is useful for feeding data into GPUs while preventing I/O
  bottlenecks. For durable storage, you can attach Hyperdisk
  volumes.

  12,000 GiB of Local SSD is automatically added to A4 instances.
  For workloads that require durable block storage, you can also attach up to
  512 TiB of Hyperdisk to A4 instances. For more information
  about disk types, see [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **Dense allocation and topology aware scheduling support**: When you provision
  A4 instances, you can request blocks of densely allocated capacity. Your host
  machines are allocated physically close to each other, provisioned as blocks
  of resources, and are interconnected with a dynamic ML network fabric to
  minimize network hops and optimize for the lowest latency. Additionally, you
  can get topology information at node and cluster level that can be used for
  job placement.

### Pricing and consumption options for A4 instances

A4 instances support the following consumption and discount options:

- [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
- [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- [Future reservations in calendar mode](https://docs.cloud.google.com/compute/docs/instances/future-reservations-calendar-mode-overview.md.txt)
- [Future reservations for capacity blocks](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt)
- Resource-based committed use discounts (CUDs)

Google Cloud bills A4 instances for their attached GPUs, predefined vCPUs,
memory, and bundled Local SSD disks. Hyperdisk and network usage
are billed separately from machine type pricing. For more information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning),
[Accelerator-optimized machine type family pricing](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized),
and [GPU pricing](https://docs.cloud.google.com/compute/gpus-pricing).

### A4 machine type (NVIDIA B200)


[A4 accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4-vms)

machine types have
[NVIDIA B200 Blackwell GPUs](https://www.nvidia.com/en-us/data-center/b200/)
(`nvidia-b200`) attached and are ideal for foundation model
training and serving. For information about machine type limitations, see
[A4 limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a4-limitations).

> [!NOTE]
> **Note:** When provisioning A4 machine types, you must reserve capacity to create instances or clusters, use Spot VMs, use Flex-start VMs, or create a resize request in a MIG. For instructions on how to create A4 instances, see [Create an A3 Ultra or A4 instance](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-a3u-a4.md.txt#create-overview). .

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

<br />

#### A4 machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A4 limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A4 machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A4 machine type.
- You can only use an A4 machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- You can't use Persistent Disk (regional or zonal). You can only use [Google Cloud Hyperdisk](https://docs.cloud.google.com/compute/docs/disks/hyperdisks.md.txt).
- The A4 machine type is only available on the [Emerald Rapids CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- You can't change the machine type of an instance to or from A4 machine type. You must create a new instance with this machine type.
- A4 machine types don't support [sole-tenancy](https://docs.cloud.google.com/compute/docs/nodes/sole-tenant-nodes.md.txt).
- You can't run Windows operating systems on an A4 machine type.
- For A4 instances, when you use `ethtool -S` to monitor GPU networking, physical port counters that end in `_phy` don't update. This is expected behavior for instances that use the MRDMA Virtual Function (VF) architecture. For more information, see [MRDMA functions and network monitoring tools](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#mrdma-monitoring).
- You can't attach Hyperdisk ML disks that were created before February 4, 2026 to A4 machine types.

### Supported disk types for A4 instances

A4 instances can use the following block storage types:

- Hyperdisk Balanced (`hyperdisk-balanced`): this is the only disk type that is supported for the boot disk
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Hyperdisk ML (`hyperdisk-ml`)
- Local SSD: which is automatically added to instances that are created by using any of the A4 machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD |
| `a4-highgpu-8g` | 128 | 128 | N/A | 128 | 8 | 32 |
|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk usage is charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />

#### Disk and capacity limits


You can attach a mixture of different Hyperdisk types to
an instance, but the maximum total disk capacity (in TiB) across all disk
types can't exceed 512 TiB for all Hyperdisks.

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance).

<br />

<br />

### Maintenance experience for A4 instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
A4

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| A4 | 8 | Minimum of 90 days | Stop and restart (with [Local SSD data persistence](https://docs.cloud.google.com/compute/docs/instances/host-maintenance-overview.md.txt#disk-persistence)) | 90 days | Yes | No |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

For more information about host maintenance in AI Hypercomputer, see
[Understand host maintenance](https://docs.cloud.google.com/ai-hypercomputer/docs/manage/host-events.md.txt#review)
in the AI Hypercomputer documentation.

## The A3 machine series

The A3 machine series has up to 224 vCPUs, and 2,944 GB of memory. This machine
series is optimized for compute and memory intensive, network bound ML training,
and HPC workloads. The A3 machine series is available in A3 Ultra, A3 Mega,
A3 High, and A3 Edge machine types.

The following table compares the features that differ between the A3 machine
types:

| Feature | A3 Ultra (NVIDIA H200) | A3 Mega, High, and Edge (NVIDIA H100) |
| GPU acceleration | [NVIDIA H200 SXM GPUs](https://www.nvidia.com/en-us/data-center/h200/) are attached, offering 141 GB of GPU memory per GPU and providing larger and faster memory to support large language models and HPC workloads. | [NVIDIA H100 SXM GPUs](https://www.nvidia.com/en-us/data-center/h100/) are attached, offering 80 GB of GPU memory per GPU, which is ideal for large transformer-based language models, databases, and HPC. |
| Intel Xeon Scalable processors | 5th Generation Intel Xeon Scalable processor (Emerald Rapids), which offers up to 4.0 GHz sustained single-core max turbo frequency. For more information about this processor, see [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt). | 4th Generation Intel Xeon Scalable processor (Sapphire Rapids), which offers up to 3.3 GHz sustained single-core max turbo frequency. For more information about this processor, see [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt). |
| Enhanced networking | For this machine type, RDMA over Converged Ethernet (RoCE) increases the network performance by combining [NVIDIA ConnectX-7 network interface cards (NICs)](https://www.nvidia.com/content/dam/en-zz/Solutions/networking/ethernet-adapters/connectx-7-datasheet-Final.pdf) with Google's datacenter-wide four-way rail-aligned network. By using RDMA over Converged Ethernet (RoCE), the `a3-ultragpu-8g` machine type achieves much higher throughput between instances in a cluster when compared to other A3 machine types. > [!NOTE] > **Note:** Because of the difference in network topology between A3 Ultra and the previous A3 series (A3 Mega, High, and Edge), you can't move workloads between instances that run on A3 Ultra and the previous A3 series. | - For the *A3 Mega* machine types, GPUDirect-TCPXO further improves on GPUDirect-TCPX by offloading the TCP protocol. By using GPUDirect-TCPXO, the `a3-megagpu-8g` machine type doubles the network bandwidth when compared to the A3 High and A3 Edge machine types. - For the *A3 Edge (`a3-edgegpu-8g`) and A3 High (`a3-highgpu-8g`)* machine types, GPUDirect-TCPX increases the network performance by allowing data packet payloads to transfer directly from GPU memory to the network interface. By using GPUDirect-TCPX, these machine types achieve much higher throughput between instances in a cluster when compared to the A2 or G2 accelerator-optimized machine types. |
| Improved networking speeds | Offers up to 4x networking speeds when compared to the previous generation A2 machine series. For more information about networking, see [Network bandwidths and GPUs](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#a3-bandwidth). | Offers up to 2.5x networking speeds when compared to the previous generation A2 machine series. For more information about networking, see [Network bandwidths and GPUs](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#a3-bandwidth). |
|---|---|---|

All A3 machine types share the following features:

- **NVLink scalability** : NVIDIA H200 and H100 GPUs provide
  peak
  [GPU NVLink bandwidth](https://www.nvidia.com/en-us/data-center/nvlink/)
  of 450 GBps, unidirectionally. By using an all-to-all NVLink topology, an
  8-GPU system achieves an aggregate bandwidth of up to 7.2 TBps.

- **Virtualization optimizations**: The Peripheral Component Interconnect
  Express (PCIe) topology of A3 instances provides more accurate locality
  information that workloads can use to optimize data transfers. The GPUs also
  expose Function Level Reset (FLR) for graceful recovery from failures and
  atomic operations support for concurrency improvements in certain scenarios.

- **Disk support**: A3 instances support Local SSD for fast scratch disks, which
  is useful for feeding data into GPUs while preventing I/O bottlenecks. For
  durable storage, you can attach Persistent Disk and Hyperdisk
  volumes.

  Local SSD is attached as follows:
  - 12,000 GiB of Local SSD is automatically added to A3 Ultra instances.
  - 6,000 GiB of Local SSD is automatically added to A3 Mega, High, and Edge instances.

  For workloads that require durable block storage, you can also attach up to
  512 TiB of Persistent Disk and Hyperdisk to machine types in
  these series. For select machine types, up to 257 TiB of Persistent Disk is
  also supported. For more information about disk types, see
  [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **Compact placement policy support** : Provides you with more control over the
  physical placement of your instances within data centers. This placement
  enables lower latency and higher bandwidth for instances that are located
  within a single availability zone. For more information, see
  [About compact placement policies](https://docs.cloud.google.com/compute/docs/instances/placement-policies-overview.md.txt#about-compact-policies).

> [!CAUTION]
> **Caution:** By default, you can't apply compact placement policies with a max distance value to A3 VMs in Compute Engine. To request access to this feature, contact your assigned [Technical Account Manager (TAM)](https://cloud.google.com/tam) or the [Sales team](https://cloud.google.com/contact).

### Pricing and consumption options for A3 instances

A3 instances support the following consumption and discount options:

- **A3 Ultra** :
  - [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
  - [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
  - [Future reservations in calendar mode](https://docs.cloud.google.com/compute/docs/instances/future-reservations-calendar-mode-overview.md.txt)
  - [Future reservations for capacity blocks](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt)
  - Resource-based committed use discounts (CUDs)
- **A3 Mega, A3 High (8 GPUs), and A3 Edge** :
  - On-demand
  - [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
  - [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
  - [On-demand reservations](https://docs.cloud.google.com/compute/docs/instances/reservations-overview.md.txt)
  - [Future reservations](https://docs.cloud.google.com/compute/docs/instances/future-reservations-overview.md.txt)
  - [Future reservations in calendar mode](https://docs.cloud.google.com/compute/docs/instances/future-reservations-calendar-mode-overview.md.txt)
  - [Future reservations for capacity blocks](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt)
  - Resource-based CUDs
- **A3 High (1, 2, or 4 GPUs)** :
  - [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
  - [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)

Google Cloud bills A3 instances for their attached GPUs, predefined vCPUs,
memory, and bundled Local SSD disks. Persistent Disk, Hyperdisk, and
network usage are billed separately from machine type pricing. For more
information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning),
[Accelerator-optimized machine type family pricing](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized),
and [GPU pricing](https://docs.cloud.google.com/compute/gpus-pricing).

### A3 Ultra machine type (NVIDIA H200)


[A3 Ultra](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-ultra-vms)
machine types have [NVIDIA H200 SXM GPUs](https://www.nvidia.com/en-us/data-center/h200/)
(`nvidia-h200-141gb`) attached and provide the highest network
performance in the A3 series. A3 Ultra machine types are ideal for foundation model training and
serving. For information about machine type limitations, see
[A3 Ultra limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-ultra-limitations).

> [!NOTE]
> **Note:** When provisioning A3 Ultra machine types, you must reserve capacity to create instances or clusters, use Spot VMs, use Flex-start VMs, or create a resize request in a MIG. For more information about the parameters to set when creating an A3 Ultra instance, see [Create an A3 Ultra or A4 instance](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-a3u-a4.md.txt#create-overview).

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

<br />

#### A3 Ultra machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A3 Ultra limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A3 Ultra machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A3 Ultra machine type.
- You can only use an A3 Ultra machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- You can't use Persistent Disk (regional or zonal). You can only use [Google Cloud Hyperdisk](https://docs.cloud.google.com/compute/docs/disks/hyperdisks.md.txt).
- The A3 Ultra machine type is only available on the [Emerald Rapids CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- Machine type changes aren't supported for A3 Ultra machine type. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on an A3 Ultra machine type.
- A3 Ultra machine types don't support [sole-tenancy](https://docs.cloud.google.com/compute/docs/nodes/sole-tenant-nodes.md.txt).
- For A3 Ultra instances, when you use `ethtool -S` to monitor GPU networking, physical port counters that end in `_phy` don't update. This is expected behavior for instances that use the MRDMA Virtual Function (VF) architecture. For more information, see [MRDMA functions and network monitoring tools](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt#mrdma-monitoring).

### A3 Mega machine type (NVIDIA H100)

[A3 Mega](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-mega-vms) machine types have [NVIDIA H100 SXM GPUs](https://www.nvidia.com/en-us/data-center/h100/) and are ideal for large model training and multi-host inference. For information about machine type limitations, see [A3 Mega limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-mega-limitations).

> [!NOTE]
> **Note:** When provisioning `a3-megagpu-8g` machine types, we recommend using a cluster of these instances and deploying with a scheduler such as Google Kubernetes Engine (GKE) or Slurm. For detailed instructions on either of these options, review the following:
>
> - To create Google Kubernetes Engine cluster, see [Deploy an A3 Mega cluster
>   with GKE](https://docs.cloud.google.com/kubernetes-engine/docs/how-to/gpu-bandwidth-gpudirect-tcpx).
> - To create a Slurm cluster, see [Deploy an A3 Mega Slurm cluster](https://docs.cloud.google.com/cluster-toolkit/docs/deploy/slurm/deploy-a3-mega-cluster.md.txt).

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

<br />

#### A3 Mega machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A3 Mega limitations (NVIDIA H100)

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A3 Mega machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A3 Mega machine type.
- You can only use an A3 Mega machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- You can't use [regional
  Persistent Disk](https://docs.cloud.google.com/compute/docs/disks/regional-persistent-disk) on an instance that uses an A3 Mega machine type.
- The A3 Mega machine type is only available on the [Sapphire Rapids CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- Machine type changes aren't supported for A3 Mega machine type. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on an A3 Mega machine type.

### A3 High machine type (NVIDIA H100)

[A3 High](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-high-vms) machine types have [NVIDIA H100 SXM GPUs](https://www.nvidia.com/en-us/data-center/h100/) and are well-suited for both large model inference and model fine tuning. For information about machine type limitations, see [A3 High limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-high-limitations).

> [!NOTE]
> **Note:** When provisioning `a3-highgpu-1g`, `a3-highgpu-2g`, or `a3-highgpu-4g` machine types, you must create instances by using Spot VMs or Flex-start VMs. For detailed instructions on these options, review the following:
>
> - To create Spot VMs, set the provisioning model to `SPOT` when you [create an accelerator-optimized
>   VM](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-accelerator-optimized).
> - To create Flex-start VMs, you can use one of the following methods:
>   - Create a standalone VM and set the provisioning model to `FLEX_START` when you [create an
>     accelerator-optimized VM](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-accelerator-optimized).
>   - Create a resize request in a managed instance group (MIG). For instructions, see [Create a MIG with GPU
>     VMs](https://docs.cloud.google.com/compute/docs/instance-groups/create-mig-with-gpu-vms).

| Machine type | vCPU count | Instance memory (GB) | Attached Local SSD (GiB) | Physical NIC count | Maximum network bandwidth (Gbps) | NVIDIA H100 GPU count | GPU memory (GB HBM3) |
| `a3-highgpu-1g` | 26 | 234 | 750 | 1 | 25 | 1 | 80 |
| `a3-highgpu-2g` | 52 | 468 | 1,500 | 1 | 50 | 2 | 160 |
| `a3-highgpu-4g` | 104 | 936 | 3,000 | 1 | 100 | 4 | 320 |
| `a3-highgpu-8g` | 208 | 1,872 | 6,000 | 5 | 1,000 | 8 | 640 |
|---|---|---|---|---|---|---|---|

<br />

#### A3 High machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A3 High limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A3 High machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A3 High machine type.
- You can only use an A3 High machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- You can't use [regional
  Persistent Disk](https://docs.cloud.google.com/compute/docs/disks/regional-persistent-disk) on an instance that uses an A3 High machine type.
- The A3 High machine type is only available on the [Sapphire Rapids CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- Machine type changes aren't supported for A3 High machine type. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on an A3 High machine type.
- For `a3-highgpu-1g`, `a3-highgpu-2g`, and`a3-highgpu-4g` machine types, you must create instances by using Spot VMs or Flex-start VMs. For detailed instructions on these options, review the following:
  - To create Spot VMs, set the provisioning model to `SPOT` when you [create an accelerator-optimized
    VM](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-accelerator-optimized).
  - To create Flex-start VMs, you can use one of the following methods:
    - Create a standalone VM and set the provisioning model to `FLEX_START` when you [create an
      accelerator-optimized VM](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-accelerator-optimized).
    - Create a resize request in a managed instance group (MIG). For instructions, see [Create a MIG with GPU
      VMs](https://docs.cloud.google.com/compute/docs/instance-groups/create-mig-with-gpu-vms).
- You can only use a [Confidential VM](https://docs.cloud.google.com/confidential-computing/confidential-vm/docs/confidential-vm-overview.md.txt) with an `a3-highgpu-1g` machine type in limited [regions and zones](https://docs.cloud.google.com/confidential-computing/confidential-vm/docs/supported-configurations.md.txt#nvidia-confidential-computing_1), and all the [limitations for Confidential VM running on the A3 High machine type](https://docs.cloud.google.com/confidential-computing/confidential-vm/docs/supported-configurations.md.txt#nvidia-confidential-computing) apply.

### A3 Edge machine type (NVIDIA H100)

[A3 Edge](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-edge-vms) machine types have [NVIDIA H100 SXM GPUs](https://www.nvidia.com/en-us/data-center/h100/) and are designed specifically for serving and are available in a [limited set of regions](https://docs.cloud.google.com/compute/docs/gpus/gpu-regions-zones#view-using-table). For information about machine type limitations, see [A3 Edge limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a3-edge-limitations).

> [!NOTE]
> **Note:** To get started with A3 Edge instances, see [Create an A3 VM with GPUDirect-TCPX enabled](https://docs.cloud.google.com/compute/docs/gpus/gpudirect.md.txt).

| Machine type | vCPU count | Instance memory (GB) | Attached Local SSD (GiB) | Physical NIC count | Maximum network bandwidth (Gbps) | NVIDIA H100 GPU count | GPU memory (GB HBM3) |
| `a3-edgegpu-8g` | 208 | 1,872 | 6,000 | 5 | 600 (`asia-south1` and `northamerica-northeast2`) or 400 (all other [A3 Edge regions](https://docs.cloud.google.com/compute/docs/gpus/gpu-regions-zones#view-using-table)) | 8 | 640 |
|---|---|---|---|---|---|---|---|

<br />

#### A3 Edge machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A3 Edge limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A3 Edge machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A3 Edge machine type.
- You can only use an A3 Edge machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- You can't use [regional
  Persistent Disk](https://docs.cloud.google.com/compute/docs/disks/regional-persistent-disk) on an instance that uses an A3 Edge machine type.
- The A3 Edge machine type is only available on the [Sapphire Rapids CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- Machine type changes aren't supported for A3 Edge machine type. To switch to or from this machine type, you must create a new instance.
- You can't run Windows operating systems on an A3 Edge machine type.
- A3 Edge machine types don't support [sole-tenancy](https://docs.cloud.google.com/compute/docs/nodes/sole-tenant-nodes.md.txt).

### Supported disk types for A3 instances

### A3 Ultra

A3 Ultra instances can use the following block storage types:

- Hyperdisk Balanced (`hyperdisk-balanced`): this is the only disk type that is supported for the boot disk
- Hyperdisk Balanced High Availability (`hyperdisk-balanced-high-availability`)
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Hyperdisk ML (`hyperdisk-ml`)
- Local SSD: which is automatically added to instances that are created by using any of the A3 machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD disks |
| `a3-ultragpu-8g` | 128 | 128 | 128 | N/A | 128 | 8 | 32 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk usage is charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

### A3 Mega

A3 Mega instances can use the following block storage types:

- Balanced Persistent Disk (`pd-balanced`)
- SSD (performance) Persistent Disk (`pd-ssd`)
- Hyperdisk Balanced (`hyperdisk-balanced`)
- Hyperdisk Balanced High Availability (`hyperdisk-balanced-high-availability`)
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Local SSD: which is automatically added to instances that are created by using any of the A3 machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD disks |
| `a3-megagpu-8g` | 128 | 32 | 32 | 64 | 64 | 8 | 16 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk and Persistent Disk usage are charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Persistent Disk and Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

### A3 High

A3 High instances can use the following block storage types:

- Balanced Persistent Disk (`pd-balanced`)
- SSD (performance) Persistent Disk (`pd-ssd`)
- Hyperdisk Balanced (`hyperdisk-balanced`)
- Hyperdisk Balanced High Availability (`hyperdisk-balanced-high-availability`)
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Local SSD: which is automatically added to instances that are created by using any of the A3 machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD disks |
| `a3-highgpu-1g` | 128 | 32 | 32 | 64 | 64 | N/A | 2 |
| `a3-highgpu-2g` | 128 | 32 | 32 | 64 | 64 | N/A | 4 |
| `a3-highgpu-4g` | 128 | 32 | 32 | 64 | 64 | 8 | 8 |
| `a3-highgpu-8g` | 128 | 32 | 32 | 64 | 64 | 8 | 16 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk and Persistent Disk usage are charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Persistent Disk and Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

### A3 Edge

A3 Edge instances can use the following block storage types:

- Balanced Persistent Disk (`pd-balanced`)
- SSD (performance) Persistent Disk (`pd-ssd`)
- Hyperdisk Balanced (`hyperdisk-balanced`)
- Hyperdisk Balanced High Availability (`hyperdisk-balanced-high-availability`)
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Extreme (`hyperdisk-extreme`)
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Local SSD: which is automatically added to instances that are created by using any of the A3 machine types

The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Throughput | Hyperdisk ML | Hyperdisk Extreme | Attached Local SSD |
| `a3-edgegpu-8g` | 128 | 32 | 32 | 64 | 64 | 8 | 16 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk and Persistent Disk usage are charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Persistent Disk and Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />

#### Disk and capacity limits


If supported by the machine type, you can attach a mixture of
Hyperdisk and Persistent Disk volumes to an instance, but the following
restrictions apply:

- The combined number of both Hyperdisk and Persistent Disk volumes can't exceed 128 per instance.
- The maximum total disk capacity (in TiB) across all disk types can't
  exceed:

  - For machine types with less than 32 vCPUs:

    - 257 TiB for all Hyperdisk or all Persistent Disk
    - 257 TiB for a mixture of Hyperdisk and Persistent Disk
  - For machine types with 32 or more vCPUs:

    - 512 TiB for all Hyperdisk
    - 512 TiB for a mixture of Hyperdisk and Persistent Disk
    - 257 TiB for all Persistent Disk

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance) and
[Persistent Disk
maximum capacity](https://docs.cloud.google.com/compute/docs/disks/persistent-disks#capacity_257tb).

<br />

### Maintenance experience for A3 instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
A3

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| A3 Ultra | 8 | Minimum of 90 days | Stop and restart (with [Local SSD data persistence](https://docs.cloud.google.com/compute/docs/instances/host-maintenance-overview.md.txt#disk-persistence)) | 90 days | Yes | No |
| A3 Mega and A3 High | 8 | - **On-demand**: minimum of 30 days - **Future reservations for capacity blocks**: minimum of 90 days | - **On-demand**: stops and restarts - **Future reservations for capacity blocks** : stops and restarts (with [Local SSD data persistence](https://docs.cloud.google.com/compute/docs/instances/host-maintenance-overview.md.txt#disk-persistence)) | - **On-demand**: 7 days - **Future reservations for capacity blocks**: 90 days | Yes | Yes |
| A3 High | 1, 2, 4 | Minimum of 30 days | Stop and restart | 7 days | No | Yes |
| A3 Edge | 8 | Minimum of 30 days | Stop and restart | 7 days | Yes | Yes |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

For more information about host maintenance in AI Hypercomputer, see
[Understand host maintenance](https://docs.cloud.google.com/ai-hypercomputer/docs/manage/host-events.md.txt#review)
in the AI Hypercomputer documentation.

## The A2 machine series

The A2 machine series is available in A2 Standard and A2 Ultra machine types.
These machine types have 12 to 96 vCPUs, and up to 1,360 GB of memory.

VM instances created by using the A2 machine types provide the following
features:

- **GPU acceleration** : each A2 instance has
  [NVIDIA A100 GPUs](https://www.nvidia.com/en-us/data-center/a100/).
  These are available in both A100 40GB and A100 80GB options.

- **NVLink scalability** : NVIDIA A100 GPUs provide peak
  [GPU to GPU NVLink bandwidth](https://www.nvidia.com/en-us/design-visualization/nvlink-bridges/)
  of 600 GBps.
  For example, systems with 16 GPUs have an aggregate NVLink bandwidth of up
  to 9.6 TBps. These 16 GPUs can be used as a single high performance
  accelerator with unified memory space to deliver up to 10 petaFLOPS of
  compute power and up to 20 petaFLOPS of inference compute power that can be
  used for artificial intelligence, deep learning, and machine learning workloads.

- **Improved computing speeds**: the attached NVIDIA A100 GPUs offer up
  to 10x improvements in computing speed when compared to previous generation
  NVIDIA V100 GPUs.

  With the A2 machine series, you can get up to 100 Gbps network bandwidth.
- **Disk support**: A2 instances support Local SSD for fast scratch disks,
  which is useful for feeding data into GPUs while preventing I/O
  bottlenecks. For durable storage, you can attach Persistent Disk and
  Hyperdisk volumes.

  Local SSD is supported as follows:
  - For A2 Standard machine types, you can add up to 3,000 GiB of Local SSD when you create an instance.
  - For A2 Ultra machine types, Local SSD is automatically attached when you create an instance.

  For workloads that require durable block storage, you can attach
  up to 257 TiB of Persistent Disk and 512 TiB of Hyperdisk
  volumes to A2 instances. For more information about disk types, see
  [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **Compact placement policy support** : provides you with more control over the
  physical placement of your instances within data centers. This enables
  lower-latency and higher bandwidth for instances that are located within a
  single availability zone. For more information, see
  [Reduce latency by using compact placement policies](https://docs.cloud.google.com/compute/docs/instances/use-compact-placement-policies.md.txt).

### Pricing and consumption options for A2 instances

A2 instances support the following consumption and discount options:

- On-demand
- [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
- [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- [On-demand reservations](https://docs.cloud.google.com/compute/docs/instances/reservations-overview.md.txt)
- [Future reservations](https://docs.cloud.google.com/compute/docs/instances/future-reservations-overview.md.txt)
- Resource-based committed use discounts (CUDs)

Google Cloud bills A2 instances for their attached GPUs, predefined vCPUs,
memory, and bundled Local SSD disks (for A2 Ultra). Persistent Disk,
Hyperdisk, additional Local SSD disks (for A2 Standard), and
network usage are billed separately from machine type pricing. For more
information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning),
[Accelerator-optimized machine type family pricing](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized),
and [GPU pricing](https://docs.cloud.google.com/compute/gpus-pricing).

The following machine types are available for the A2 machine series.

### A2 Ultra machine types (NVIDIA A100 80GB)

These machine types have a
[fixed number of A100 80GB GPUs](https://docs.cloud.google.com/compute/docs/gpus.md.txt#a100-gpus).
Local SSD is automatically attached to instances created by using the A2 Ultra
machine types. For information about machine type limitations, see
[A2 Ultra limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a2-ultra-limitations).

| Machine type | vCPU count | Instance memory (GB) | Attached Local SSD (GiB) | Maximum network bandwidth (Gbps) | NVIDIA A100 80GB GPU count | GPU memory (GB HBM2e) |
| `a2-ultragpu-1g` | 12 | 170 | 375 | 24 | 1 | 80 |
| `a2-ultragpu-2g` | 24 | 340 | 750 | 32 | 2 | 160 |
| `a2-ultragpu-4g` | 48 | 680 | 1,500 | 50 | 4 | 320 |
| `a2-ultragpu-8g` | 96 | 1,360 | 3,000 | 100 | 8 | 640 |
|---|---|---|---|---|---|---|

<br />

#### A2 Ultra machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A2 Ultra limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A2 Ultra machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A2 Ultra machine type.
- You can only use an A2 Ultra machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- The A2 Ultra machine type is only available on the [Cascade Lake platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- If your instance uses an A2 Ultra machine type, you can't change the machine type. If you need to use a different A2 Ultra machine type, or any other machine type, you must create a new instance.
- You can't change any other machine type to an A2 Ultra machine type. If you need a instance that uses an A2 Ultra machine type, you must create a new instance.
- You can't do a quick format of the attached Local SSDs on Windows instances that use A2 Ultra machine types. To format these Local SSDs, you must do a full format by using the [diskpart
  utility](https://docs.microsoft.com/windows-server/administration/windows-commands/diskpart) and specifying `format fs=ntfs label=tmpfs`.

### A2 Standard machine types (NVIDIA A100 40GB)

These machine types have a [fixed number of A100 40GB GPUs](https://docs.cloud.google.com/compute/docs/gpus.md.txt#a100-gpus).
You can also add Local SSD disks when creating an A2 Standard instance. For the
number of disks you can attach, see
[Machine types that require you to choose a number of Local SSD disks](https://docs.cloud.google.com/compute/docs/disks/local-ssd.md.txt#lssd_disk_options).
For information about machine type limitations, see
[A2 Standard limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#a2-standard-limitations).

| Machine type | vCPU count | Instance memory (GB) | Local SSD supported | Maximum network bandwidth (Gbps) | NVIDIA A100 40GB GPU count | GPU memory (GB HBM2) |
| `a2-highgpu-1g` | 12 | 85 | Yes | 24 | 1 | 40 |
| `a2-highgpu-2g` | 24 | 170 | Yes | 32 | 2 | 80 |
| `a2-highgpu-4g` | 48 | 340 | Yes | 50 | 4 | 160 |
| `a2-highgpu-8g` | 96 | 680 | Yes | 100 | 8 | 320 |
| `a2-megagpu-16g` | 96 | 1,360 | Yes | 100 | 16 | 640 |
|---|---|---|---|---|---|---|

<br />

#### A2 Standard machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### A2 Standard limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for an A2 Standard machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use an A2 Standard machine type.
- You can only use an A2 Standard machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- The A2 Standard machine type is only available on the [Cascade Lake platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- If your instance uses an A2 Standard machine type, you can only switch from one A2 Standard machine type type to another A2 Standard machine type. You can't change to any other machine type. For more information, see [Modify accelerator-optimized instances](https://docs.cloud.google.com/compute/docs/gpus/add-remove-gpus.md.txt#accelerator-optimized_vms).
- You can't use the Windows operating system with the `a2-megagpu-16g` machine type. When using a Windows operating system, choose a different A2 Standard machine type.
- You can't do a quick format of the attached Local SSDs on Windows instances that use A2 Standard machine types. To format these Local SSDs, you must do a full format by using the [diskpart
  utility](https://docs.microsoft.com/windows-server/administration/windows-commands/diskpart) and specifying `format fs=ntfs label=tmpfs`.

### Supported disk types for A2 instances

A2 instances can use the following block storage types:

- Hyperdisk ML (`hyperdisk-ml`)
- Balanced Persistent Disk (`pd-balanced`)
- SSD (performance) Persistent Disk (`pd-ssd`)
- Standard Persistent Disk (`pd-standard`)
- Local SSD: which is automatically attached to instances created by using the A2 Ultra machine types.


### A2 Ultra

The following table lists the maximum number of disks supported per instance:

| Machine types | All disks | Hyperdisk ML | Attached Local SSD |
| `a2-ultragpu-1g` | 128 | 32 | 1 |
| `a2-ultragpu-2g` | 128 | 48 | 2 |
| `a2-ultragpu-4g` | 128 | 64 | 4 |
| `a2-ultragpu-8g` | 128 | 64 | 8 |
|---|---|---|---|

The following notes apply to the preceding table:

- **All disks**: this limit applies to Persistent Disk and Hyperdisk, and doesn't include Local SSD disks.
- **Disk pricing** : Hyperdisk and Persistent Disk usage are charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Persistent Disk and Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

### A2 Standard

The following table lists the maximum number of disks supported per instance:

| Machine types | All disks | Hyperdisk ML | Local SSD |
| `a2-highgpu-1g` | 128 | 32 | 8 |
| `a2-highgpu-2g` | 128 | 48 | 8 |
| `a2-highgpu-4g` | 128 | 64 | 8 |
| `a2-highgpu-8g` | 128 | 64 | 8 |
| `a2-megagpu-16g` | 128 | 64 | 8 |
|---|---|---|---|

The following notes apply to the preceding table:

- **All disks**: this limit applies to Persistent Disk and Hyperdisk, and doesn't include Local SSD disks.
- **Disk pricing** : Hyperdisk and Persistent Disk usage are charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Persistent Disk and Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />


If supported by the machine type, you can attach a mixture of
Hyperdisk and Persistent Disk volumes to an instance, but the following
restrictions apply:

- The combined number of both Hyperdisk and Persistent Disk volumes can't exceed 128 per instance.
- The maximum total disk capacity (in TiB) across all disk types can't
  exceed:

  - For machine types with less than 32 vCPUs:

    - 257 TiB for all Hyperdisk or all Persistent Disk
    - 257 TiB for a mixture of Hyperdisk and Persistent Disk
  - For machine types with 32 or more vCPUs:

    - 512 TiB for all Hyperdisk
    - 512 TiB for a mixture of Hyperdisk and Persistent Disk
    - 257 TiB for all Persistent Disk

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance) and
[Persistent Disk
maximum capacity](https://docs.cloud.google.com/compute/docs/disks/persistent-disks#capacity_257tb).

<br />

### Maintenance experience for A2 instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
A2

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| A2 Ultra | 1, 2, 4, 8 | Minimum of 30 days | Stop and restart | 7 days | Yes (8 GPUs only) | Yes |
| A2 Standard | 1, 2, 4, 8, or 16 | Minimum of 30 days | Stop and restart | 7 days | Yes (8 and 16 GPUs only) | Yes |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

## The G4 machine series

The G4 machine series uses the AMD EPYC Turin CPU platform and features
NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. This machine series offers
significant improvements over the previous-generation G2 machine series,
with considerably more GPU memory, increased GPU memory bandwidth, and
higher networking bandwidth.

G4 instances have up to 384 vCPUs, 1,440 GB of memory, and 12 TiB of
Titanium SSD disks attached. G4 instances also provide up to
400 Gbps of standard network performance.

This machine series is particularly intended for workloads such as NVIDIA
Omniverse simulation workloads, graphics-intensive applications, video transcoding, and
virtual desktops. The G4 machine series also provide a low-cost solution for
performing single host inference and model tuning compared with A series machine
types.

Instances that use the G4 machine type provide the following features:

- **GPU acceleration with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs** :
  G4 instances automatically attach
  [NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs](https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/),
  which offer 96 GB GPU memory per GPU.

- **5th Generation AMD EPYC Turin CPU Platform** : this platform offers up to
  4.1 GHz sustained max boost frequency. For more information about this
  processor, see [CPU platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).

- **Next generation graphics performance**: the NVIDIA RTX PRO 6000 GPUs provide
  significant performance and feature upgrades over the NVIDIA L4 GPUs that
  are attached to the G2 machine series. Thesed upgrades are as follows:

  - **5th-Generation Tensor Cores**: these cores introduce support for FP4 precision and DLSS 4 Multi Frame Generation. By using these 5th Generation tensor cores, NVIDIA RTX PRO 6000 GPUs offers improved performance to accelerate tasks like local LLM development and content creation, compared to NVIDIA L4 GPUs.
  - **4th-Generation RT Cores**: these cores deliver up to twice the ray-tracing performance of the previous generation NVIDIA L4 GPUs, accelerating rendering for design and manufacturing workloads.
  - **Core count**: the NVIDIA RTX PRO 6000 GPU includes 24,064 CUDA cores, 752 5th-gen Tensor cores, and 188 4th-gen RT cores. This update represents a substantial increase over prior generations like the L4 GPU, which has 7,680 CUDA cores and 240 Tensor cores.
- **GPU sharing**: there are a number of options that you can use to allow
  multiple workloads to access a single physical GPU. GPU sharing is useful
  for workloads that don't require the resources of a full GPU, helping you
  optimize costs. The following GPU sharing options are available for G4
  instances:

  - **Fractional GPU (vGPU) support** : this feature allows a single physical GPU to be shared by multiple virtual machine (VM) instances. vGPUs provide multi-tenant security isolation because each vGPU is a separate VM instance. To use vGPUs, Compute Engine provides the following VM instance shapes: `g4-standard-6` (1/8 GPU), `g4-standard-12` (1/4 GPU), and `g4-standard-24` (1/2 GPU).
  - **Multi-Instance GPU (MIG)**: this feature allows a single GPU to be partitioned into up to four fully isolated instances on a single virtual machine. This option doesn't offer multi-tenant security isolation as all partitions belong to a single VM.
- **Peripheral Component Interconnect Express (PCIe) Gen 5 support**: G4
  instances supports PCI Express Gen 5, which improves the data transfer speed
  from CPU memory to GPU compared to PCIe Gen 3 used by G2 instances.

- **Disk support** : G4 instances support
  [Titanium SSD](https://docs.cloud.google.com/compute/docs/disks/local-ssd.md.txt#local_ssd_types) for
  fast scratch disks, which is useful for feeding data into GPUs while
  preventing I/O bottlenecks. For durable storage, you can attach
  Hyperdisk volumes.

  G4 instances support attaching up to 12,000 GiB of Titanium SSD.
  For workloads that require durable block storage, G4 instances also
  support attaching up to 512 TiB of Hyperdisk. For more
  information about disk types, see [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **GPU Peer-to-Peer (P2P) communication** : G4 instances support GPU P2P
  communication, enabling direct data transfer between GPUs within the same
  instance. This can significantly improve performance for multi-GPU workloads
  by reducing data transfer latency and freeing up CPU resources. For more
  information, see [G4 GPU peer-to-peer (P2P) communication](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-gpu-p2p).

- **Compact placement policy support** : this feature gives you control over the
  physical location of your instances within data centers. Compact placement
  policies help you collocate instances close together to reduce network
  latency, which is useful for latency-sensitive workloads like high performance
  computing (HPC) or machine learning (ML). For more information, see
  [Reduce latency by using compact placement policies](https://docs.cloud.google.com/compute/docs/instances/use-compact-placement-policies.md.txt).

### Pricing and consumption options for G4 instances

G4 instances support the following consumption and discount options:

- On-demand
- [Compute flexible committed use discounts (CUDs)](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based)
- Resource-based committed use discounts (CUDs)
- [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
- [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- [On-demand reservations](https://docs.cloud.google.com/compute/docs/instances/reservations-overview.md.txt)
- [Future reservations](https://docs.cloud.google.com/compute/docs/instances/future-reservations-overview.md.txt)

Google Cloud bills G4 instances for their attached GPUs, predefined vCPUs,
and memory. Hyperdisk, Titanium SSD disks, and network usage
are billed separately from machine type pricing. For more information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning),
[Accelerator-optimized machine type family pricing](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized),
and [GPU pricing](https://docs.cloud.google.com/compute/gpus-pricing).

### G4 machine types (NVIDIA RTX PRO 6000)


[G4 accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-series)

machine types use [NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs](https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/) (`nvidia-rtx-pro-6000`)
and are
suitable for NVIDIA Omniverse simulation workloads, graphics-intensive applications, video
transcoding, and virtual desktops. G4 machine types also provide a low-cost solution for
performing single host inference and model tuning compared with A series machine types.

You can also add Titanium SSD disks when you create a G4 instance. For the number of disks
you can attach, see
[Machine types that require you to choose a number of Local SSD disks](https://docs.cloud.google.com/compute/docs/disks/local-ssd.md.txt#lssd_disk_options).
For information about machine type limitations, see
[G4 limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g4-limitations).

> [!NOTE]
> **Note:** To get started with G4 instances, see [Create a G4 instance](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-g-series.md.txt).

| Machine type | vCPU count | Instance memory (GB) | Maximum Titanium SSD supported (GiB) | Physical NIC count | Maximum network bandwidth (Gbps) | NVIDIA RTX PRO 6000 GPU count | GPU memory (GB GDDR7) |
| `g4-standard-6` | 6 | 22 | 0 | 1 | 20 | 1/8 | 12 |
| `g4-standard-12` | 12 | 45 | 375 | 1 | 20 | 1/4 | 24 |
| `g4-standard-24` | 24 | 90 | 750 | 1 | 20 | 1/2 | 48 |
| `g4-standard-48` | 48 | 180 | 1,500 | 1 | 50 | 1 | 96 |
| `g4-standard-96` | 96 | 360 | 3,000 | 1 | 100 | 2 | 192 |
| `g4-standard-192` | 192 | 720 | 6,000 | 1 | 200 | 4 | 384 |
| `g4-standard-384` | 384 | 1,440 | 12,000 | 2 | 400 | 8 | 768 |
|---|---|---|---|---|---|---|---|

<br />

#### G4 machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### G4 limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for a G4 machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use a G4 machine type.
- You can only use a G4 machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- You can't use [Persistent Disk](https://docs.cloud.google.com/compute/docs/disks/persistent-disks.md.txt) (regional or zonal) on an instance that uses a G4 machine type.
- The G4 machine type is only available on the [AMD EPYC Turin 5th Generation platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#amd_processors).
- You can only use a [Confidential VM](https://docs.cloud.google.com/confidential-computing/confidential-vm/docs/confidential-vm-overview.md.txt) with the `g4-standard-48` machine type in limited [regions and zones](https://docs.cloud.google.com/confidential-computing/confidential-vm/docs/supported-configurations.md.txt#nvidia-confidential-computing_1). All the [limitations for Confidential VM running on the G4 machine type](https://docs.cloud.google.com/confidential-computing/confidential-vm/docs/supported-configurations.md.txt#nvidia-confidential-computing) apply.
- You can't use Windows operating systems on `g4-standard-384` instances.
- You can't attach Hyperdisk ML disks that were created before February 4, 2026 to G4 machine types.
- When creating G4 instances that have less than one GPU attached (fractional GPUs), don't use the `--no-service-account` or `--no-scopes` flags. To authenticate NVIDIA vGPU drivers, Compute Engine must [verify the VM's identity](https://docs.cloud.google.com/compute/docs/instances/verifying-instance-identity.md.txt). This process requires that service accounts are enabled.

### Supported disk types for G4 instances

G4 instances can use the following block storage types:

- Hyperdisk Balanced (`hyperdisk-balanced`): this is the only disk type that is supported for the boot disk
- Hyperdisk Balanced High Availability (`hyperdisk-balanced-high-availability`)
- Hyperdisk Extreme (`hyperdisk-extreme`): this disk type is only supported on G4 instances that have two or more GPUs attached
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Titanium SSD: you can add Titanium SSD to instances created by using the G4 machine types.


The following table lists the maximum number of disks supported per instance:

| Machine types | All Hyperdisk | Hyperdisk Balanced | Hyperdisk Balanced High Availability | Hyperdisk Extreme | Hyperdisk ML | Hyperdisk Throughput | Titanium SSD |
| `g4-standard-6` | 8 | 8 | 8 | 0 | 8 | 8 | 0 |
| `g4-standard-12` | 16 | 16 | 16 | 0 | 16 | 16 | 1 |
| `g4-standard-24` | 32 | 32 | 32 | 0 | 32 | 32 | 2 |
| `g4-standard-48` | 32 | 32 | 32 | 0 | 32 | 32 | 4 |
| `g4-standard-96` | 32 | 32 | 32 | 8 | 32 | 32 | 8 |
| `g4-standard-192` | 64 | 64 | 64 | 8 | 64 | 64 | 16 |
| `g4-standard-384` | 128 | 128 | 128 | 8 | 128 | 128 | 32 |
|---|---|---|---|---|---|---|---|

The following note applies to the preceding table:

- **Disk pricing** : Hyperdisk usage is charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />


You can attach a mixture of different Hyperdisk types to
an instance, but the maximum total disk capacity (in TiB) across all disk
types can't exceed 512 TiB for all Hyperdisks.

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance).

<br />

<br />

### GPU peer-to-peer (P2P) communication

G4 instances enhance multi-GPU workload performance by using direct GPU
peer-to-peer (P2P) communication, which is supported only on machine types with
two or more GPUs. This approach allows GPUs that attach to the same G4 instance
to exchange data directly over the PCIe bus, bypassing the need to transfer data
through the CPU's main memory. This direct path reduces latency, lowers CPU
utilization, and increases the effective bandwidth between GPUs. P2P
communication significantly accelerates multi-GPU applications such as machine
learning (ML) training and high performance computing (HPC).

Because G4 instances use GPU P2P communication by default, you receive the
following benefits without modifying your application code. However, for
applications that use the NVIDIA Collective Communication Library (NCCL), to
receive the full performance improvements from P2P communication, you must
[configure NCCL to use P2P](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#configure-nccl-p2p).

#### Key benefits and performance

- **Accelerates multi-GPU workloads on G4 instances with two or more GPUs** : provides faster runtimes for applications running on `g4-standard-96`, `g4-standard-192`, and `g4-standard-384` machine types.
- **Provides high-bandwidth communication**: enables high data transfer speeds between GPUs.
- **Improves NCCL performance**: provides significant performance improvements for
  applications that use the NVIDIA Collective Communication Library (NCCL)
  when compared to communication that doesn't use P2P. Google's hypervisor
  securely isolates this P2P communication within your instances.

  - On four GPU instances (`g4-standard-192`), all GPUs are on a single NUMA node, allowing for the most efficient P2P communication. This can lead to performance improvements of up to *2.04x* for collectives such as `Allgather`, `Allreduce`, and `ReduceScatter`.
  - On eight GPU instances (`g4-standard-384`), GPUs are distributed across two NUMA nodes. P2P communication is accelerated for traffic both within and between these nodes, with performance improvements of up to *2.19x* for the same collectives.

#### Configure NCCL to use P2P

To configure NCCL to use P2P, before you run your workloads, set the
[`NCCL_P2P_LEVEL` environment variable](https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html#nccl-p2p-level)
on your G4 instance based on the machine type:

- For G4 instances with 2 or 4 GPUs (`g4-standard-96`, `g4-standard-192`): set `NCCL_P2P_LEVEL=PHB`
- For G4 instances with 8 GPUs (`g4-standard-384`): set `NCCL_P2P_LEVEL=SYS`

Set the environment variable by using one of the following options:

- On the command line, run the appropriate export command, such as `export NCCL_P2P_LEVEL=SYS`, in the shell session where you plan to run your app. To make this setting persistent, add this command to your shell's startup script, such as `~/.bashrc`.
- Add the appropriate setting, such as `NCCL_P2P_LEVEL=SYS`, to the NCCL configuration file located at `/etc/nccl.conf`.

### Maintenance experience for G4 instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
G4

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| G4 | 1, 2, or 4 | Minimum of 30 days | Stop and restart. If Local SSD disks are attached, the instance stops with Local SSD data persistence. | 7 days | No | Yes |
| G4 | 8 | Minimum of 90 days | Stop and restart. If Local SSD disks are attached, the instance stops with Local SSD data persistence. | 30 days | Yes | Yes |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

## The G2 machine series

The G2 machine series is available in standard machine types that have 4 to 96
vCPUs, and up to 432 GB of memory. This machine series is optimized for
inference and graphics workloads. The G2 machine series is available in a single
standard machine type with multiple configurations.

Instances created by using the G2 machine types provide the following features:

- **GPU acceleration** : each G2 machine type has
  [NVIDIA L4 GPUs](https://www.nvidia.com/en-us/data-center/l4/).

- **Improved inference rates** : the G2 machine type provides support for the
  [FP8 (8-bit floating point)](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/examples/fp8_primer.html)
  data type which speeds up ML inference rates and
  reduces memory requirements.

- **Next generation graphics performance** : NVIDIA L4 GPUs provide
  up to 3X improvement in graphics performance by using third-generation
  [RT cores](https://www.nvidia.com/en-us/geforce/ada-lovelace-architecture/#rt-cores)
  and
  [NVIDIA DLSS 3 (Deep Learning Super Sampling)](https://www.nvidia.com/en-us/geforce/technologies/dlss/)
  technology.

- **High performance network bandwidth**: with the G2 machine types, you can
  get up to 100 Gbps network bandwidth.

- **Disk support**: G2 instances support Local SSD for fast scratch disks,
  which is useful for feeding data into GPUs while preventing I/O
  bottlenecks. For durable storage, you can attach Persistent Disk and
  Hyperdisk volumes.

  You can add up to 3,000 GiB of Local SSD to G2 instances.
  For workloads that require durable block storage, you can attach
  Hyperdisk and Persistent Disk volumes to
  G2 instances. The maximum storage capacity depends on the number of vCPUs
  the instance has. For more information about disk types, see
  [Choose a disk type](https://docs.cloud.google.com/compute/docs/disks.md.txt).
- **Compact placement policy support** : provides you with more control over the
  physical placement of your instances within data centers. This enables
  lower-latency and higher bandwidth for instances that are located within a
  single availability zone. For more information, see
  [Reduce latency by using compact placement policies](https://docs.cloud.google.com/compute/docs/instances/use-compact-placement-policies.md.txt).

### Pricing and consumption options for G2 instances

G2 instances support the following consumption and discount options:

- On-demand
- [Compute flexible committed use discounts (CUDs)](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based)
- Resource-based committed use discounts (CUDs)
- [Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt)
- [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt)
- [On-demand reservations](https://docs.cloud.google.com/compute/docs/instances/reservations-overview.md.txt)
- [Future reservations](https://docs.cloud.google.com/compute/docs/instances/future-reservations-overview.md.txt)

Google Cloud bills G2 instances for their attached GPUs, predefined vCPUs,
and memory. Persistent Disk, Hyperdisk, Local SSD disks, and network
usage are billed separately from machine type pricing. For more information, see
[Pricing and consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning),
[Accelerator-optimized machine type family pricing](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized),
and [GPU pricing](https://docs.cloud.google.com/compute/gpus-pricing).

### G2 machine types (NVIDIA L4)

[G2 accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-vms)
machine types have [NVIDIA L4 GPUs](https://www.nvidia.com/en-us/data-center/l4/)
attached and are ideal for cost-optimized inference, graphics-intensive and
high performance computing workloads.

Each G2 machine type also has a default memory and a custom
memory range. The custom memory range defines the amount of memory that
you can allocate to your instance for each machine type. You can also add Local
SSD disks when creating a G2 instance. For the number of disks
you can attach, see
[Machine types that require you to choose a number of Local SSD disks](https://docs.cloud.google.com/compute/docs/disks/local-ssd.md.txt#lssd_disk_options).

For information about machine type limitations, see
[G2 limitations](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-limitations).

| Machine type | vCPU count | Default instance memory (GB) | Custom instance memory range (GB) | Max Local SSD supported (GiB) | Maximum network bandwidth (Gbps) | NVIDIA L4 GPU count | GPU memory (GB GDDR6) |
| `g2-standard-4` | 4 | 16 | 16 to 32 | 375 | 10 | 1 | 24 |
| `g2-standard-8` | 8 | 32 | 32 to 54 | 375 | 16 | 1 | 24 |
| `g2-standard-12` | 12 | 48 | 48 to 54 | 375 | 16 | 1 | 24 |
| `g2-standard-16` | 16 | 64 | 54 to 64 | 375 | 32 | 1 | 24 |
| `g2-standard-24` | 24 | 96 | 96 to 108 | 750 | 32 | 2 | 48 |
| `g2-standard-32` | 32 | 128 | 96 to 128 | 375 | 32 | 1 | 24 |
| `g2-standard-48` | 48 | 192 | 192 to 216 | 1,500 | 50 | 4 | 96 |
| `g2-standard-96` | 96 | 384 | 384 to 432 | 3,000 | 100 | 8 | 192 |
|---|---|---|---|---|---|---|---|

<br />

#### G2 machine type notes


The following notes apply to the preceding table:

- **vCPU count** : a vCPU is implemented as a single hardware hyper-thread on one of the available [CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt).
- **Maximum network bandwidth** : maximum egress bandwidth can't 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](https://docs.cloud.google.com/compute/docs/network-bandwidth.md.txt).
- **GPU memory**: GPU 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.

<br />

#### G2 limitations

- You can only request capacity by using the [supported consumption options](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#pricing-and-provisioning) for a G2 machine type.
- You don't receive [sustained
  use discounts](https://docs.cloud.google.com/compute/docs/sustained-use-discounts) and [flexible committed use](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based) discounts for instances that use a G2 machine type.
- You can only use a G2 machine type in certain [regions
  and zones](https://docs.cloud.google.com/compute/docs/regions-zones/gpu-regions-zones#view-using-table).
- The G2 machine type is only available on the [Cascade Lake platform](https://docs.cloud.google.com/compute/docs/cpu-platforms.md.txt#intel_processors).
- Standard Persistent Disk (`pd-standard`) isn't supported on instances that use the G2 machine type. For supported disk types, see [Supported disk types for G2](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines.md.txt#g2-disks).
- You can't create [Multi-Instance
  GPUs](https://www.nvidia.com/en-us/technologies/multi-instance-gpu/) on an instance that uses a G2 machine type.
- If you need to change the machine type of a G2 instance, review [Modify accelerator-optmized instances](https://docs.cloud.google.com/compute/docs/gpus/add-remove-gpus.md.txt#accelerator-optimized_vms).
- You can't use [Deep Learning VM Images](https://docs.cloud.google.com/deep-learning-vm/docs/images.md.txt) as boot disks for instances that use the G2 machine type.
- You must use the Google Cloud CLI or REST to [create G2 instances](https://docs.cloud.google.com/compute/docs/gpus/create-gpu-vm-accelerator-optimized.md.txt#create-vm) for the following scenarios:
  - You want to specify custom memory values.
  - You want to customize the number of visible CPU cores.

### Supported disk types for G2 instances

G2 instances can use the following block storage types:

- Balanced Persistent Disk (`pd-balanced`)
- SSD (performance) Persistent Disk (`pd-ssd`)
- Hyperdisk ML (`hyperdisk-ml`)
- Hyperdisk Throughput (`hyperdisk-throughput`)
- Local SSD: you can add Local SSD to instances created by using the G2 machine types.


The following table lists the maximum number of disks supported per instance:

| Machine types | All disks | Hyperdisk ML | Hyperdisk Throughput | Local SSD |
| `g2-standard-4` | 128 | 24 | 24 | 1 |
| `g2-standard-8` | 128 | 32 | 32 | 1 |
| `g2-standard-12` | 128 | 32 | 32 | 1 |
| `g2-standard-16` | 128 | 48 | 48 | 1 |
| `g2-standard-24` | 128 | 48 | 48 | 2 |
| `g2-standard-32` | 128 | 64 | 64 | 1 |
| `g2-standard-48` | 128 | 64 | 64 | 4 |
| `g2-standard-96` | 128 | 64 | 64 | 8 |
|---|---|---|---|---|

The following notes apply to the preceding table:

- **All disks**: this limit applies to Persistent Disk and Hyperdisk, and doesn't include Local SSD disks.
- **Disk pricing** : Hyperdisk and Persistent Disk usage are charged separately from [machine type pricing](https://cloud.google.com/compute/vm-instance-pricing). For disk pricing, see [Persistent Disk and Hyperdisk pricing](https://docs.cloud.google.com/compute/disks-image-pricing#persistentdisk).

<br />


If supported by the machine type, you can attach a mixture of
Hyperdisk and Persistent Disk volumes to an instance, but the following
restrictions apply:

- The combined number of both Hyperdisk and Persistent Disk volumes can't exceed 128 per instance.
- The maximum total disk capacity (in TiB) across all disk types can't
  exceed:

  - For machine types with less than 32 vCPUs:

    - 257 TiB for all Hyperdisk or all Persistent Disk
    - 257 TiB for a mixture of Hyperdisk and Persistent Disk
  - For machine types with 32 or more vCPUs:

    - 512 TiB for all Hyperdisk
    - 512 TiB for a mixture of Hyperdisk and Persistent Disk
    - 257 TiB for all Persistent Disk

For details about the capacity limits, see
[Hyperdisk size and attachment limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits.md.txt#limits-instance) and
[Persistent Disk
maximum capacity](https://docs.cloud.google.com/compute/docs/disks/persistent-disks#capacity_257tb).

<br />

### Maintenance experience for G2 instances


During the [lifecycle of a
Compute Engine instance](https://docs.cloud.google.com/compute/docs/instances/instance-lifecycle), the host machine that your instance runs on undergoes multiple
*host events*.

A host event can include the regular maintenance of
Compute Engine infrastructure, or in rare cases, a host error. Compute Engine also
applies some non-disruptive lightweight upgrades for the hypervisor and network
in the background.

<br />

The following table describes the host maintenance features for
G2

machine types:

| Machine type | Number of GPUs | Typical scheduled maintenance event frequency | [Maintenance behavior](https://docs.cloud.google.com/compute/docs/instances/setting-vm-host-options.md.txt#available_host_maintenance_properties) | Advanced notification for scheduled maintenance | On-demand maintenance | Simulate maintenance |
| G2 | 1, 2, 4, or 8 | Minimum of 30 days | Stop and restart | 7 days | Yes (8 GPUs only) | Yes |
|---|---|---|---|---|---|---|

The maintenance frequencies shown in the previous table are approximations, not guarantees.
Compute Engine might occasionally perform maintenance more frequently.

## Pricing and consumption options

Consumption options refers to the ways to get and use compute resources.
Google Cloud bills accelerator-optimized machine types for their attached GPUs,
predefined vCPU, memory, and bundled Local SSD (if applicable). Discounts
for accelerator-optimized instances can vary depending on the resource type and
consumption option that you use. At any given point, a resource is eligible for
only one type of discount.

Depending on the resource type and consumption option that you choose, discounts
apply to accelerator-optimized machine types as follows:

- **On-demand:** With this consumption option, you can receive
  [compute flexible committed use discounts (CUDs)](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview.md.txt#spend_based)
  only for eligible spend on N1, G2, and G4 instance usage, as follows:

  - For N1 machine types that have attached GPUs, you can receive compute flexible CUDs only for vCPUs, memory, and Local SSD disks.
  - For G2 and G4 machine types, you can receive compute flexible CUDs for vCPUs, memory, GPUs, and Local SSD disks.

  To receive resource-based CUDs for other accelerator-optimized machine
  series, including G2 and G4, use one of the reservation options instead.
- **Spot:** Spot VMs automatically receive discounts through
  [Spot VMs pricing](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt#pricing).

- **Flex-start:** Instances provisioned by using the Flex-start consumption
  option automatically receive discounts through
  [Dynamic Workload Scheduler pricing](https://cloud.google.com/products/dws/pricing).

- **Reservations:** You can receive resource-based CUDs for your
  accelerator-optimized machine type resources by purchasing resource-based
  commitments. Resource-based commitments for all GPUs and most Local SSD
  disks
  [require attached reservations for those resources](https://docs.cloud.google.com/compute/docs/instances/signing-up-committed-use-discounts.md.txt#commitments_for_gpus_and_local_ssd_disk_resources).

For more pricing information for accelerator-optimized instances, see the
[Accelerator-optimized machine type family](https://docs.cloud.google.com/compute/vm-instance-pricing#accelerator-optimized)
section on the VM instance pricing page.

### Consumption option availability by machine type

The following table summarizes the availability of each consumption option by
machine types. For more information about how to choose a consumption option,
see [Choose a consumption model](https://docs.cloud.google.com/ai-hypercomputer/docs/consumption-models.md.txt#choose)
in the AI Hypercomputer documentation.

> [!NOTE]
> **Note:** Before you create and submit a future reservation request for a supported GPU machine type, you must contact your [account team](https://cloud.google.com/tam) or the [sales team](https://cloud.google.com/contact) to discuss your request. Otherwise, Google Cloud is likely to decline it.

| Machine type (GPU model) | On-demand | [Spot](https://docs.cloud.google.com/compute/docs/instances/spot.md.txt) | [Flex-start](https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms.md.txt) | [On-demand reservations](https://docs.cloud.google.com/compute/docs/instances/reservations-overview.md.txt) | [Future reservations](https://docs.cloud.google.com/compute/docs/instances/future-reservations-overview.md.txt) | [Future reservations in calendar mode](https://docs.cloud.google.com/compute/docs/instances/future-reservations-calendar-mode-overview.md.txt) | [Future reservations for capacity blocks](https://docs.cloud.google.com/ai-hypercomputer/docs/reserve-capacity.md.txt) |
| A5X (Vera Rubin) | No | No | No | No | No | No | Yes |
| A4X Max (GB300) and A4X (GB200) | No | No | No | No | No | No | Yes |
| A4 (B200) | No | Yes | Yes | No | No | Yes | Yes |
| A3 Ultra (H200) | No | Yes | Yes | No | No | Yes | Yes |
| A3 Mega (H100) | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| A3 High with 8 GPUs (H100) | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| A3 High with less than 8 GPUs (H100) | No | Yes | Yes | No | No | No | No |
| A3 Edge (H100) | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
| A2 (A100) | Yes | Yes | Yes | Yes | Yes | No | No |
| G4 (RTX PRO 6000) | Yes | Yes | Yes | Yes | Yes | No | No |
| G2 (L4) | Yes | Yes | Yes | Yes | Yes | No | No |
| N1 with GPUs (T4/P4/V100) | Yes | Yes | Yes | Yes | Yes | No | No |
|---|---|---|---|---|---|---|---|

## What's next

- [Create a VM with attached GPUs](https://docs.cloud.google.com/compute/docs/gpus/create-vm-with-gpus.md.txt)
- [GPU pricing](https://cloud.google.com/compute/gpus-pricing)
- [VM instance pricing](https://cloud.google.com/compute/vm-instance-pricing)
- [Networking and GPU machines](https://docs.cloud.google.com/compute/docs/gpus/gpu-network-bandwidth.md.txt)
- [AI Hypercomputer overview](https://docs.cloud.google.com/ai-hypercomputer/docs/overview.md.txt)