Accelerator-optimized machine family


The accelerator-optimized machine family is designed by Google Cloud to deliver the needed performance and efficiency for GPU accelerated workloads such as artificial intelligence (AI), machine learning (ML), and high performance computing (HPC).

The accelerator-optimized machine family is available in the following machine series: A3, A2 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 to N1 general-purpose machine types.

Machine series recommendation by workload type

The following section provide the recommended machine series based on your GPU workloads.

Large AI models

Workload type Best fit Good alternative
Multiple (distributed) server training A3 A2
Inference A3, A2

Mainstream models

Workload type Best fit Good alternative (in recommended order)
Multiple (distributed) server training A3
  • A2
  • G2
  • N1+V100
Single server training A3, A2
  • G2
  • N1+V100
Inference G2
  • N1+T4
  • N1+V100

Graphics-intensive workloads

Workload type Best fit (in recommended order)
Video streaming and transcoding, remote virtual workstations, digital twins
  • G2
  • N1+T4

High performance computing

For high performance computing workloads, any accelerator-optimized machine series works well. The best fit depends on the amount of computation that must be offloaded to the GPU.

Pricing and discount

All accelerator-optimized machine types support the following discount and consumption options:

The accelerator-optimized machine types are billed for their attached GPUs, predefined vCPU, memory, and bundled Local SSD (if applicable). For more pricing information for accelerator-optimized VMs, see Accelerator-optimized machine type family section on the VM instance pricing page.

The A3 machine series

The A3 machine series is available in a standard machine type and has 208 vCPUs, and up to 1872 GB of memory. This machine series is optimized for compute and memory intensive, network bound ML training, and HPC workloads.

The A3 machine series also provides the following features:

  • Next generation hardware: each A3 machine type has NVIDIA H100 GPUs attached, which offers 80GB GPU memory per GPU and is ideal for large transformer-based language models, databases, and HPC.

    This machine series is built on the 4th Generation Intel Xeon Scalable processor (Sapphire Rapids) and offers up to 3.3 GHz sustained single-core max turbo frequency.

  • Industry-leading NVLink scale: NVIDIA H100 GPUs provide peak GPU NVLink bandwidth of 450 GB/s, unidirectionally. With all-to-all NVLink topology between 8 GPUs in a system, the aggregate NVLink Bandwidth is up to 7.2 TB/s. These GPUs can be used as a single high performance accelerator with unified memory space to deliver up to 25 petaFLOPS of AI/DL/ML compute power and up to 50 petaFLOPS of inference compute power.

  • Improved computing speed and networking: NVIDIA H100 GPUs offer up to 2.5X improvement in computing speed and 10x the virtual machine (VM) network bandwidth when compared to previous generation NVIDIA A100 GPUs.

    Each a3-highgpu-8g VM has five physical network interface cards (NICs), four of which share the same Peripheral Component Interconnect Express (PCIe) bus and have a non-uniform memory access (NUMA) alignment of one NIC per two NVIDIA H100 80GB GPUs. These four physical NICs are ideal for dedicated high bandwidth GPU to GPU communication. The other physical NIC resides on a separate PCIe bus and is ideal for other networking needs. Each NIC has a maximum bandwidth of 200 Gbps, which totals to a maximum bandwidth of 1000 Gbps (1 Tbps) per VM.
  • Improved GPU cluster performance with GPUDirect-TCPX: GPUDirect-TCPX increases the network performance by allowing data packet payloads to transfer directly from GPU memory to the network interface.

    By leveraging GPUDirect-TCPX, A3 VMs deliver the highest throughput between VMs in a cluster when compared to the A2 or G2 accelerator-optimized machine types.

    To maximize GPU cluster performance with GPUDirect-TCPX, review the following documents:

  • Virtualization optimizations: The Peripheral Component Interconnect Express (PCIe) topology of A3 VMs provides more accurate locality information that workloads can use to optimize data transfers.

    The NVIDIA H100 GPUs also expose Function Level Reset (FLR) for graceful recovery from failures and atomic operations support for concurrency improvements in certain scenarios.

  • Storage: 6 TB of Local SSD is automatically added to VMs created by using the A3 machine type. Local SSD can be used for fast scratch disks or for feeding data into the GPUs while preventing I/O bottlenecks.

    You can also attach up to 257 TB of Persistent Disk storage to the machine types in these series for applications that require higher storage performance.

  • Compact placement policy support: provides you with more control over the physical placement of your VMs within data centers. This enables lower-latency and higher bandwidth for VM placement within a single availability zone. Compact placement policy supports up to 96 VMs in a lower-latency subset of the network, within a given zone.

Supported disk types for A3

A3 VMs can use the following block storage types:

  • Balanced Persistent Disk (pd-balanced)
  • SSD (performance) Persistent Disk (pd-ssd)
  • Hyperdisk Extreme (hyperdisk-extreme)
  • Hyperdisk Throughput (hyperdisk-throughput)
  • Local SSD: which is automatically added to VMs that are created by using the A3 machine type

Disk and capacity limits

You can use a mixture of Persistent Disk and Hyperdisk volumes with a VM, but the following restrictions apply:

  • The combined number of both Hyperdisk and Persistent Disk volumes can't exceed 128 per VM.
  • 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 Hyperdisk
      • 257 TiB for a mixture of Hyperdisk and Persistent Disk
      • 257 TiB for Persistent Disk
    • For machine types with 32 or more vCPUs:

      • 512 TiB for Hyperdisk
      • 512 TiB for a mixture of Hyperdisk and Persistent Disk
      • 257 TiB for Persistent Disk

For more details about these limits, see Hyperdisk capacity limits per VM and Persistent Disk maximum capacity.

A3 standard machine types

This machine type has a fixed number of H100 GPUs.

Machine types GPU count vCPUs* Memory (GB) Max number of disks per VM,
across all disks
Max total Hyperdisk
volumes per VM
Max total disk size
across all disks (TiB)#
Local SSD Maximum egress bandwidth (Gbps)
a3-highgpu-8g 8 208 1872 128
  • Hyperdisk Throughput—64
  • Hyperdisk Extreme—8
512, including up to 257 TiB of Persistent Disk Yes 1000

*A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
Hyperdisk and Persistent Disk usage are charged separately from machine type pricing.
Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. See Network bandwidth.
#For Hyperdisk Throughput, the maximum size per volume is 32 TiB. For all other Persistent Disk or Hyperdisk volumes, the maximum size is 64 TiB.

A3 standard limitations

  • You don't receive sustained use discounts and flexible committed use discounts for VMs that use A3 standard machine types.
  • You can only use A3 standard machine types in certain regions and zones.
  • You can't use regional persistent disks on VMs that use A3 standard machine types.
  • The A3 standard machine type is only available on the Sapphire Rapids platform.
  • If your VM uses an A3 standard machine type, you can't change the machine type. If you need to use another machine type, you must create a new VM.
  • You can't change any other machine type to an A3 standard machine type. If you need to create a VM that uses an A3 standard machine type, you must create a new VM.
  • A3 standard machine types don't support sole-tenancy.
  • You can't run A3 standard machine types on Windows operating systems.

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 1360 GB of memory.

The A2 machine series also provides the following features:

  • NVIDIA GPUs attached: each A2 machine type has NVIDIA A100 GPUs. These are available in both A100 40GB and A100 80GB options.

  • Industry-leading NVLink scale that provides peak GPU to GPU NVLink bandwidth 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 speed: 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.

  • Storage: for fast scratch disks or for feeding data into the GPUs while preventing I/O bottlenecks, the A2 machine types support Local SSD as follows:

    • For the A2 standard machine types, you can add up to 3 TB of Local SSD.
    • For the A2 ultra machine types, Local SSD is automatically attached when you create the VM.

    You can also attach up to 257 TB of Persistent Disk storage to A2 VMs for applications that require this higher storage performance.

  • Compact placement policy support: provides you with more control over the physical placement of your VMs within data centers. This enables lower-latency and higher bandwidth for VM placement within a single availability zone. For more information, see Reduce latency by using compact placement policies.

Supported disk types for A2

A2 VMs can use the following block storage types:

  • Balanced Persistent Disk (pd-balanced)
  • SSD (performance) Persistent Disk (pd-ssd)
  • Standard Persistent Disk (pd-standard)
  • Local SSD: which is automatically attached to VMs created by using the A2 ultra machine types.

A2 standard machine types

These machine types have a fixed number of A100 40GB GPUs.

Machine types GPU count vCPUs* Memory (GB) Max number of Persistent Disk volumes Max total Persistent Disk (TB) Local SSD Maximum egress bandwidth (Gbps)
a2-highgpu-1g 1 12 85 128 257 Yes 24
a2-highgpu-2g 2 24 170 128 257 Yes 32
a2-highgpu-4g 4 48 340 128 257 Yes 50
a2-highgpu-8g 8 96 680 128 257 Yes 100
a2-megagpu-16g 16 96 1360 128 257 Yes 100

*A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
Hyperdisk and Persistent Disk usage are charged separately from machine type pricing.
Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. See Network bandwidth.
#For Hyperdisk Throughput, the maximum size per volume is 32 TiB. For all other Persistent Disk or Hyperdisk volumes, the maximum size is 64 TiB.

A2 standard limitations

  • You don't receive sustained use discounts and flexible committed use discounts for VMs that use A2 standard machine types.
  • You can only use A2 standard machine types in certain regions and zones.
  • You can't use regional persistent disks on VMs that use A2 standard machine types.
  • The A2 standard machine type is only available on the Cascade Lake platform.
  • If your VM uses an A2 standard machine type, you can only switch from one A2 standard machine type to another A2 standard machine type. You can't change to any other machine type. For more information, see Modify accelerator-optmized VMs.
  • You can't use the a2-megagpu-16g A2 standard machine type on Windows operating systems. When using Windows operating systems, choose a different A2 standard machine type.
  • You can't do a quick format of the attached Local SSDs on Windows VMs that use A2 standard machine types. To format these Local SSDs, you must do a full format by using the diskpart utility and specifying format fs=ntfs label=tmpfs.
  • A2 standard machine types don't support sole-tenancy.

A2 ultra machine types

These machine types have a fixed number of A100 80GB GPUs. Local SSD is automatically attached to VMs created by using the A2 ultra machine types.

Machine types GPU count vCPUs* Memory (GB) Max number of Persistent Disk volumes Max total Persistent Disk (TB) Bundled Local SSD Maximum egress bandwidth (Gbps)
a2-ultragpu-1g 1 12 170 128 257 375 GB 24
a2-ultragpu-2g 2 24 340 128 257 750 GB 32
a2-ultragpu-4g 4 48 680 128 257 1.5 TB 50
a2-ultragpu-8g 8 96 1360 128 257 3 TB 100

*A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
Hyperdisk and Persistent Disk usage are charged separately from machine type pricing.
Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. See Network bandwidth.
#For Hyperdisk Throughput, the maximum size per volume is 32 TiB. For all other Persistent Disk or Hyperdisk volumes, the maximum size is 64 TiB.

A2 ultra limitations

  • You don't receive sustained use discounts and flexible committed use discounts for VMs that use A2 ultra machine types.
  • You can only use A2 ultra machine types in certain regions and zones.
  • You can't use regional persistent disks on VMs that use A2 ultra machine types.
  • The A2 ultra machine type is only available on the Cascade Lake platform.
  • If your VM 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 VM.
  • You can't change any other machine type to an A2 ultra machine type. If you need to create a VM that uses an A2 ultra machine type, you must create a new VM.
  • You can't do a quick format of the attached Local SSDs on Windows VMs that use A2 ultra machine types. To format these Local SSDs, you must do a full format by using the diskpart utility and specifying format fs=ntfs label=tmpfs.

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 also provides the following features:

  • NVIDIA GPUs attached: each G2 machine type has NVIDIA L4 GPUs.

  • Improved inference rates: the G2 machine types provide support for the FP8 (8-bit floating point) 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 and NVIDIA DLSS 3 (Deep Learning Super Sampling) technology.

  • High performance network bandwidth: with the G2 machine series, you can get up to 100 Gbps network bandwidth.

  • Storage: you can add up to 3 TB of Local SSD to G2 VMs. This can be used for fast scratch disks or for feeding data into the GPUs while preventing I/O bottlenecks.

    You can also attach Hyperdisk and Persistent Disk volumes to G2 VMs, for applications that require more persistent storage. The maximum storage capacity depends on the number of vCPUs the VM has. For details, see Supported disk types.

  • Compact placement policy support: provides you with more control over the physical placement of your VMs within data centers. This enables lower-latency and higher bandwidth for VM placement within a single availability zone. For more information, see Reduce latency by using compact placement policies.

Supported disk types for G2

G2 VMs can use the following block storage types:

  • Balanced Persistent Disk (pd-balanced)
  • SSD (performance) Persistent Disk (pd-ssd)
  • Hyperdisk Throughput (hyperdisk-throughput)
  • Local SSD

You can use a mixture of Persistent Disk and Hyperdisk volumes with a VM, but the following restrictions apply:

  • The combined number of both Hyperdisk and Persistent Disk volumes can't exceed 128 per VM.
  • 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 Hyperdisk
      • 257 TiB for a mixture of Hyperdisk and Persistent Disk
      • 257 TiB for Persistent Disk
    • For machine types with 32 or more vCPUs:

      • 512 TiB for Hyperdisk
      • 512 TiB for a mixture of Hyperdisk and Persistent Disk
      • 257 TiB for Persistent Disk

For more details about these limits, see Hyperdisk capacity limits per VM and Persistent Disk maximum capacity.

G2 standard machine types

Each G2 machine type has a fixed number of NVIDIA L4 GPUs and vCPUs attached. 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 VM for each machine type. You can specify your custom memory during VM creation.

Machine types GPU count vCPUs* Default memory (GB) Custom memory range (GB) Max number of disks per VM,
across all disks#
Max total Hyperdisk
volumes per VM
Max total disk size
across all disks (TiB)
Max Local SSD supported (GB) Maximum egress bandwidth (Gbps)
g2-standard-4 1 4 16 16 - 32 128 24 257 375 10
g2-standard-8 1 8 32 32 - 54 128 32 257 375 16
g2-standard-12 1 12 48 48 - 54 128 32 257 375 16
g2-standard-16 1 16 64 54 - 64 128 48 257 375 32
g2-standard-24 2 24 96 96 - 108 128 64 257 750 32
g2-standard-32 1 32 128 96 - 128 128 64 512, but at most 257 TiB of Persistent Disk 375 32
g2-standard-48 4 48 192 192 - 216 128 64 512, but at most 257 TiB of Persistent Disk 1500 50
g2-standard-96 8 96 384 384 - 432 128 64 512, but at most 257 TiB of Persistent Disk 3000 100

*A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
Hyperdisk and Persistent Disk usage are charged separately from machine type pricing.
Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. See Network bandwidth.
#For Hyperdisk Throughput, the maximum size per volume is 32 TiB. For all other Persistent Disk or Hyperdisk volumes, the maximum size is 64 TiB.

G2 standard limitations

  • You don't receive sustained use discounts and flexible committed use discounts for VMs that use G2 standard machine types.
  • You can only use G2 standard machine types in certain regions and zones.
  • You can't use regional persistent disks on VMs that use G2 standard machine types.
  • The G2 standard machine type is only available on the Cascade Lake platform.
  • Standard persistent disks (pd-standard) are not supported on VMs that use G2 standard machine types. For supported disk types, see Supported disk types for G2.
  • You can't create Multi-Instance GPUs on G2 standard machine types.
  • If you need to change the machine type of a G2 VM, review Modify accelerator-optmized VMs.
  • You can't use Deep Learning VM Images as boot disks for your VMs that use G2 standard machine types.
  • The current default driver for Container-Optimized OS doesn't support L4 GPUs running on G2 machine types. Container-Optimized OS also only support a select set of drivers. If you want to use Container-Optimized OS on G2 machine types, review the following notes:
    • Use a Container-Optimized OS version that supports the minimum recommended NVIDIA driver version 525.60.13 or later. For more information, review the Container-Optimized OS release notes.
    • When you install the driver, specify the latest available version that works for the L4 GPUs. For example, sudo cos-extensions install gpu -- -version=525.60.13.
  • You must use the Google Cloud CLI or REST to create G2 VMs for the following scenarios:
    • You want to specify custom memory values.
    • You want to customize the number of visible CPU cores.

What's next