This document describes the machine families, machine series, and machine types
that you can choose from to create a virtual machine (VM) instance or bare metal
instance with the resources that you need. For accelerator-optimized machines,
this document describes only Graphics Processing Unit (GPU) accelerators.
For
information about machine types that contain Google's custom-developed Tensor
Processing Units (TPUs), see [TPU machines](https://docs.cloud.google.com/compute/docs/tpus/tpu-machines).
There are several machine families you can choose from. Each machine family is
further organized into machine series and predefined machine types within each
series. For example, within the N2 machine series in the general-purpose
machine family, you can select the `n2-standard-4` machine type.
When you create a compute instance, you select a machine type from a machine
family and series. The machine type determines the resources that
Compute Engine allocates to the instance. For example, the
`n2-standard-4` machine type creates a VM with 4 vCPUs and 16 GB of memory.
For information about machine series that support
[Spot VMs](https://docs.cloud.google.com/compute/docs/instances/spot) (and preemptible VMs), see
[Compute Engine instances provisioning models](https://docs.cloud.google.com/compute/docs/instances/provisioning-models).
**Note:** This is a list of Compute Engine machine families. For a detailed explanation of each machine family, see the following pages:
- [General-purpose](https://docs.cloud.google.com/compute/docs/general-purpose-machines)---best price-performance ratio for a variety of workloads.
- [Storage-optimized](https://docs.cloud.google.com/compute/docs/storage-optimized-machines)---best for workloads that are low in core usage and high in storage density.
- [Compute-optimized](https://docs.cloud.google.com/compute/docs/compute-optimized-machines)---designed for high performance computing (HPC) solutions and compute-intensive workloads; offers high performance per core on Compute Engine.
- [Network-optimized](https://docs.cloud.google.com/compute/docs/network-optimized-machines)---purpose-built for IO-intensive workloads that require high network bandwidth, superior packet processing performance, and superior block storage throughput and IOPS per vCPU.
- [Memory-optimized](https://docs.cloud.google.com/compute/docs/memory-optimized-machines)---ideal for memory-intensive workloads, offering more memory per core than other machine families, with up to 12 TB of memory.
- [Accelerator-optimized](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines)---ideal for massively parallelized Compute Unified Device Architecture (CUDA) compute workloads, such as machine learning (ML) and high performance computing (HPC). This family is the best option for workloads that require accelerators (GPUs or TPUs).
## Compute Engine terminology
This documentation uses the following terms:
- **Machine family**: A curated set of processor and hardware configurations
optimized for specific workloads, for example, General-purpose,
Accelerator-optimized, or Memory-optimized.
- **Machine series**: Machine families are further classified by series,
generation, and processor type.
- Each series focuses on a different aspect of computing power or
performance. For example, the E series offers efficient VMs at a low cost,
while the C series offers better performance.
- The generation is denoted by an ascending number. For example, the N1 series
within the general-purpose machine family is the older version of the N2
series.
A higher generation or series number usually indicates newer underlying
CPU platforms or technologies. For example, the M3 series, which runs on
Intel Xeon Scalable Processor 3rd Generation (Ice Lake), is a newer
generation than the M2 series, which runs on Intel Xeon Scalable Processor
2nd Generation (Cascade Lake).
| Generation | Intel | AMD | Arm |
|---|---|---|---|
| 5th generation machine series | X5 | N/A | A5X |
| 4th generation machine series | N4, C4, C4N, Z4M, X4, M4, M4N, A4 | C4D, G4, N4D, H4D, Z4D | N4A, C4A, A4X Max, A4X |
| 3rd generation machine series | C3, H3, Z3, M3, A3 | C3D | N/A |
| 2nd generation machine series | N2, E2, C2, M2, A2, G2 | N2D, C2D, T2D, E2 | T2A |
- **Machine type** : Every machine series offers at least one machine type. Each
machine type provides a set of resources for your compute instance, such as
vCPUs, memory, disks, and GPUs.
If a predefined machine type does not meet
your needs, you can also create a [custom machine type](https://docs.cloud.google.com/compute/docs/machine-resource#custom-types) for
some machine series.
The following sections describe the different machine types.
### Predefined machine types
Predefined machine types come with a non-modifiable amount of memory and
vCPUs. Predefined machine types use a variety of vCPU to memory ratios:
- `highcpu` --- from 1 to 3 GB memory per vCPU; typically, 2 GB memory per vCPU.
- `standard` --- from 3 to 7 GB memory per vCPU; typically, 4 GB memory per vCPU.
- `highmem` --- from 7 to 12 GB memory per vCPU; typically, 8 GB memory per vCPU.
- `megamem` --- from 12 to 15 GB memory per vCPU; typically, 14 GB memory per vCPU.
- `ultramem` --- from 24 to 31 GB memory per vCPU.
- `hypermem` --- from 15 to 24 GB memory per vCPU; typically,
16 GB memory per vCPU.
For example, a `c3-standard-22` machine type has 22 vCPUs, and as a
`standard` machine type, it also has 88 GB of memory.
### Local SSD machine types
Local SSD machine types are special predefined machine types. The machine type
names include `lssd`. When you create a compute instance using one of the
following machine types,
[Titanium SSD or Local SSD disks](https://docs.cloud.google.com/compute/docs/disks/local-ssd#local_ssd_types)
are automatically attached to the instance:
- `-lssd`: Available with the C4, C4A, C4D, C4N, C3, C3D, and H4D machine series, these machine types attach a predetermined number of 375 GiB Titanium SSD or Local SSD disks to the instance. Examples of this machine type include `c4a-standard-4-lssd`, `c3-standard-88-lssd`, and `c3d-highmem-360-lssd`.
- `-standardlssd`: Available with the storage-optimized Z3 and Z4D machine series, these machine types provide between 200 and 275 GiB of Titanium SSD disk capacity per vCPU. These machine types are recommended for high performance search and data analysis for medium-sized data sets. An example of this machine type is `z3-highmem-22-standardlssd`.
- `-highlssd`: Available with the Z3 and Z4D machine series, these machine types provide between 375 and 438 GiB of Titanium SSD disk capacity per vCPU. These machine types offer high performance and are recommended for storage-intensive streaming and data analysis for large-sized data sets. An example of this machine type is `z4d-highmem-96-highlssd`.
Other machine series also support Local SSD disks but don't use a machine type
name that includes `lssd`. For a list of all the machine types that you
can use with Titanium SSD or Local SSD disks, see
[Choose a valid number of Local SSD disks](https://docs.cloud.google.com/compute/docs/disks/local-ssd#choose_number_local_ssds).
### Bare metal machine types
Bare metal machine types are special predefined machine types. The machine type
name includes `-metal`. When you create a compute instance using one of these
machine types, there is no hypervisor installed on the instance. You can attach
disks to a bare metal instance, just as you would with a VM
instance. Bare metal instances can be used in VPC networks and subnetworks in
the same way as VM instances.
> [!NOTE]
> **Note:** Compute Engine bare metal instances aren't related to [Bare Metal Solution](https://docs.cloud.google.com/bare-metal/docs/bms-overview).
For more information, see
[Bare metal instances on Compute Engine](https://docs.cloud.google.com/compute/docs/instances/bare-metal-instances).
### Custom machine types
If none of the predefined machine types match your workload needs, you can
create a VM instance with a custom machine type for the
[N and E machine series](https://docs.cloud.google.com/compute/docs/general-purpose-machines#custom_machine_types)
in the general-purpose machine family.
Custom machine types cost slightly more to use compared to an equivalent
predefined machine type. Also, there are limitations in the amount of memory and
vCPUs that you can select for a custom machine type. The on-demand prices for
custom machine types include a 5% premium over the on-demand and commitment
prices for predefined machine types.
When creating a custom machine type, you can use the extended memory feature.
Instead of using the default memory size based on the number of vCPUs you
select, you can specify an amount of memory, up to the limit for the machine
series.
For more information, see
[Create a VM with a custom machine type](https://docs.cloud.google.com/compute/docs/instances/creating-instance-with-custom-machine-type).
### Shared-core machine types
The E2 and N1 series contain shared-core machine types.
These machine types timeshare a physical core which can be a
cost-effective method for running small, non-resource-intensive apps.
- [E2](https://docs.cloud.google.com/compute/docs/general-purpose-machines#sharedcore): offers
`e2-micro`, `e2-small`, and `e2-medium` shared-core machine types with
2 vCPUs for short periods of bursting.
- [N1](https://docs.cloud.google.com/compute/docs/general-purpose-machines#n1_machine_types): offers
`f1-micro` and `g1-small` shared-core machine types
which have up to 1 vCPU available for short periods of bursting.
For more information, see [CPU bursting](https://docs.cloud.google.com/compute/docs/general-purpose-machines#cpu-bursting).
## Machine family and series recommendations
The following tables provide recommendations for different workloads. To get
AI-powered assistance when you evaluate which machine family best fits your
workload requirements, compare pricing trade-offs, and generate deployment
code snippets, you can prompt Gemini in the Google Cloud console. For more
information, see
[Design compute infrastructure with Compute Advisor](https://docs.cloud.google.com/compute/docs/design-with-gemini).
| **General-purpose workloads** ||||
| N4, N4A, N4D, N2, N2D, N1 | C4, C4A, C4D, C3, C3D | E2 | Tau T2D, Tau T2A |
|---|---|---|---|
| Balanced price/performance across a wide range of machine types | Consistently high performance for a variety of workloads | Day-to-day computing at a lower cost | Best per-core performance/cost for scale-out workloads |
| - Medium traffic web and app servers - Containerized microservices - Business intelligence apps - Virtual desktops - CRM applications - Development and test environments - Batch processing - Storage and archive | - High traffic web and app servers - Databases - In-memory caches - Ad servers - Game Servers - Data analytics - Media streaming and transcoding - CPU-based ML training and inference | - Low-traffic web servers - Back office apps - Containerized microservices - Microservices - Virtual desktops - Development and test environments | - Scale-out workloads - Web serving - Containerized microservices - Media transcoding - Large-scale Java applications |
| **Optimized workloads** |||||
| **Storage-optimized** | **Compute-optimized** | **Network-optimized** | **Memory-optimized** | **Accelerator-optimized^1^ (GPUs)** |
| Z3, Z4D, and Z4M | H4D, H3, C2 and C2D | C4N and M4N | X5, X4, M4, M4N, M3, M2, M1 | A5X, A4X Max, A4X, A4, A3, A2, G4, G2 |
|---|---|---|---|---|
| Highest block storage to compute ratios for storage-intensive workloads | Highest performance and lower cost for high performance computing (HPC), multi-node and compute-bound workloads | Purpose-built for I/O-intensive workloads that require high network bandwidth, superior packet processing performance, and leading block storage throughput and IOPS per vCPU. | Highest memory to compute ratios for memory-intensive workloads | Optimized for accelerated high performance computing workloads |
| - SQL, NoSQL, and vector databases - Data analytics and data warehouses - Search - Media streaming - Large distributed parallel file systems - Checkpointing - Key-Value (KV) cache | - Manufacturing, weather forecasting, electronic design automation (EDA), High-performance web servers - Healthcare and life sciences, scientific computing - Seismic processing and structural mechanics applications - Modeling and simulation workloads, AI/ML - High-performance web servers, Game Servers | - Small to large databases (Oracle, SQL server, SAP HANA) and in-memory databases - Analytics platforms - Machine learning models - High performance file systems - Security and Network appliances | - Small to extra-large SAP HANA in-memory databases - In-memory data stores, such as Redis - Simulation - High Performance databases such as Microsoft SQL Server, MySQL - Electronic design automation | - Generative AI models such as the following: - Large Language Models (LLM) - Diffusion Models - Generative Adversarial Networks (GAN) - CUDA-enabled ML training and inference - High-performance computing (HPC) - Massively parallelized computation - BERT natural language processing - Deep learning recommendation model (DLRM) - Video transcoding - Remote visualization workstation |
After you create a compute instance, you can use *rightsizing recommendations*
to optimize resource utilization based on your workload. For more information,
see [Applying machine type recommendations for VMs](https://docs.cloud.google.com/compute/docs/instances/apply-machine-type-recommendations-for-instances).
## General-purpose machine family guide
The [**general-purpose machine family**](https://docs.cloud.google.com/compute/docs/general-purpose-machines)
offers several machine series with the best price-performance ratio for a
variety of workloads.
Compute Engine offers general-purpose machine series that run on
either x86 or Arm architecture.
### x86
- The C4 machine series is available on the Intel Granite Rapids and Emerald Rapids CPU platforms and powered by [Titanium](https://cloud.google.com/titanium). C4 machine types are optimized to deliver consistently high performance and scale up to 288 vCPUs, 2.2 TB of DDR5 memory, and 18 TiB of Local SSD. C4 is available in `highcpu` (2 GB memory per vCPU), `standard` (3.75 GB memory per vCPU), and `highmem` (7.75 GB memory per vCPU) configurations. C4 instances are aligned with the underlying [non-uniform memory access (NUMA)](https://wikipedia.org/wiki/Non-uniform_memory_access) architecture to offer optimal, reliable, and consistent performance.
- The C4D machine series is available on the AMD EPYC Turin CPU platform and powered by Titanium. C4D has a greater max boost frequency as compared with C3D, with improved Instructions Per Clock (IPC) for faster database transactions. By leveraging Hyperdisk storage and Titanium networking, C4D demonstrates up to 55% higher queries per second on Cloud SQL for MySQL and 35% better performance on Memorystore for Redis workloads as compared to C3D. C4D instances are available with up to 384 vCPUs, 3 TB of DDR5 memory, and 12 TiB of Local SSD. C4D is available in `highcpu` (1.875 GB memory per vCPU), `standard` (3.875 GB memory per vCPU), and `highmem` (7.875 GB memory per vCPU) configurations. C4D instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance.
- The N4 machine series is available on the Intel Emerald Rapids CPU platform and powered by Titanium. N4 machine types are optimized for flexibility and cost with both predefined and custom shapes and can scale up to 80 vCPUs at 640 GB of DDR5 memory. N4 is available in `highcpu` (2 GB per vCPU), `standard` (4 GB per vCPU), and `highmem` (8 GB per vCPU) configurations.
- The N4D machine series is available on the AMD EPYC Turin CPU platform and powered by Titanium. N4D machine types are built for flexibility and cost optimization through an efficient architecture, and next-generation [dynamic resource management](https://docs.cloud.google.com/compute/docs/dynamic-resource-management), making better use of resources on host machines. You can create N4D VMs using predefined machine types with up to 96 vCPUs and 768 GB of DDR5 memory, or you can create N4D VMs using custom machine types that allow you to choose varied combinations of compute and memory to optimize costs and reduce resource waste. N4D is available in `highcpu` (2 GB per vCPU), `standard` (4 GB per vCPU), and `highmem` (8 GB per vCPU) configurations.
- The N2 machine series has up to 128 vCPUs, 8 GB of memory per vCPU, and is available on the Intel Ice Lake and Intel Cascade Lake CPU platforms.
- The N2D machine series has up to 224 vCPUs, 8 GB of memory per vCPU, and is available on the third-generation AMD EPYC Milan platform.
- The C3 machine series offers up to 176 vCPUs and 2, 4, or 8 GB of memory per vCPU on the Intel Sapphire Rapids CPU platform and Titanium. C3 instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance.
- The C3D machine series offers up to 360 vCPUs and 2, 4, or 8 GB of memory per vCPU on the AMD EPYC Genoa CPU platform and Titanium. C3D instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance.
- The E2 machine series has up to 32 virtual cores (vCPUs) with up to 128 GB of memory for predefined machine types or up to 256 GB of memory for custom machine types (with a maximum of 8 GB per vCPU), and the lowest cost of all machine series. The E2 machine series has a predefined CPU platform, running either an Intel processor or an AMD processor. The processor is selected for you when you create the instance. This machine series provides a variety of compute resources for the lowest price on Compute Engine, especially when paired with [committed use discounts](https://docs.cloud.google.com/compute/docs/instances/committed-use-discounts-overview).
- The Tau T2D machine series provides an optimized feature set for scaling out. Each VM instance can have up to 60 vCPUs, 4 GB of memory per vCPU, and is available on third-generation AMD EPYC Milan processors. The Tau T2D machine series doesn't use cluster-threading, so a vCPU is equivalent to an entire core.
- The N1 machine series VMs can have up to 96 vCPUs, up to 6.5 GB of memory per vCPU, and are available on Intel Sandy Bridge, Ivy Bridge, Haswell, Broadwell, and Skylake CPU platforms.
### Arm
- The N4A machine series is powered by Google's custom-designed Axion
processor. The Axion processor is built on the Arm Neoverse N3 compute core, which
supports Arm V9.2 architecture. The N4A machine series uses
[Titanium](https://cloud.google.com/titanium)
for CPU offloading. N4A instances provide
up to 64 vCPUs with up to 8 GB of memory per vCPU
with [Uniform Memory Access (UMA) domain](https://wikipedia.org/wiki/Uniform_memory_access).
N4A instances don't use simultaneous multithreading (SMT). A vCPU in an
N4A instance is equivalent to an entire physical core.
The N4A machine series is engineered to be our most efficient and flexible
Arm-based series, delivering exceptional price-performance for a wide range
of general-purpose and scale-out workloads. Ideal use cases include web
and application servers, microservices, containerized applications using
Google Kubernetes Engine (GKE), open-source databases, and development and testing
environments.
- The C4A machine series is powered by Google Axion, and built on the Arm Neoverse
V2 compute core, which
supports Arm V9 architecture. C4A instances are powered by
Titanium with disk and network offloading, which
improves instance performance by reducing on-host processing.
C4A instances provide up to 72 vCPUs with up to 8 GB of memory per vCPU
in a single UMA domain.
C4A offers `-lssd` machine types that come with up to 6 TiB of
[Titanium SSD](https://docs.cloud.google.com/compute/docs/disks/local-ssd#local_ssd_types) capacity.
C4A bare metal instances have 96 vCPUs and up to 768 GB of memory.
C4A instances don't use simultaneous multithreading (SMT). A vCPU in a
C4A instance is equivalent to an entire physical core.
- The Tau T2A machine series is the first machine series in Google Cloud
built on Arm Neoverse N1 core compute. Tau T2A machines are optimized to
deliver compelling price for performance. Each VM can have up to 48 vCPUs
with 4 GB of memory per vCPU. The Tau T2A machine series runs on a
64 core Ampere Altra processor with an Arm instruction set and an all-core
frequency of 3 GHz. Tau T2A machine types support a single NUMA
node and a vCPU is equivalent to an entire core.
## Storage-optimized machine family guide
The [**storage-optimized machine family**](https://docs.cloud.google.com/compute/docs/storage-optimized-machines)
is best suited for high-performance and flash-optimized workloads such as SQL,
NoSQL, and vector databases, scale-out data analytics, data warehouses and
search, and distributed file systems that need fast access to large amounts of
data stored in local storage. The storage-optimized machine family is designed
to provide high local storage throughput and IOPS at sub-millisecond latency.
Z3 runs on the Intel Xeon Scalable processor (code name Sapphire Rapids) with
DDR5 memory and
[Titanium](https://cloud.google.com/titanium)
offload processors. Z3 brings together compute, networking, and storage
innovations into one platform. Z3 instances are aligned with the underlying NUMA
architecture to offer optimal, reliable, and consistent performance.
+Z3 `standardlssd` machine types can have up to 176 vCPUs,
1,406 GB of memory, and 35 TiB of Titanium SSD.
+ Z3 `highlssd` machine types can have up to 88 vCPUs, 704 GB of memory,
and 35 TiB of Titanium SSD.
+ Z3 bare metal machine types have 192 vCPUs, 1,536 GB of memory, and
70 TiB of local Titanium SSD.
Z4D runs on AMD EPYC Turin processors with DDR5 memory and Titanium
offload processors. Z4D machine types are optimized for the
underlying NUMA architecture to deliver optimal, reliable, and consistent
performance. Z4D also offers larger Titanium SSD disk sizes than Z3.
- Z4D `standardlssd` machine types can have up to 384 vCPUs, 3,024 GB of memory, and 84,000 GiB of Titanium SSD.
- Z4D `highlssd` machine types can have up to 192 vCPUs, 1,512 GB of memory, and 84,000 GiB of Titanium SSD.
- Z4D bare metal machine types have 384 vCPUs, 3,072 GB of memory, and 84 TiB of local Titanium SSD.
Z4M runs on Intel Emerald Rapids processors and Titanium offload
processors. Z4M instances provide up to 192 vCPUs, 1,488 GB of DDR5
memory, and up to 168,000 GiB of Titanium SSD. Z4M instances can scale
up to 400 Gbps networking throughput on a single Virtual Private Cloud. Disk
Titanium SSD performance can scale up to 73 GiB/s peak read,
56 GiB/s peak write, and 9.8 GiB/s sustained write to meet the storage
requirements for AI/ML training workloads.
- Z4M is available in predefined machine types in `highmem` configurations with 7.625 or 7.75 GB of memory per vCPU and 875 GiB of Titanium SSD per vCPU.
Z4M instances are aligned with the underlying NUMA architecture to offer
optimal, reliable, and consistent performance.
## Compute-optimized machine family guide
The [**compute-optimized machine family**](https://docs.cloud.google.com/compute/docs/compute-optimized-machines)
is optimized for running high performance computing (HPC), multi-node, and
compute-bound applications by providing high performance per core.
- H4D instances offer 192 vCPUs and 720 GB of DDR5 memory. H4D instances run on the AMD EPYC Turin CPU platform, with [Titanium](https://cloud.google.com/titanium) offloading and [Cloud RDMA](https://docs.cloud.google.com/compute/docs/networking/using-irdma) support. H4D instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance. H4D delivers improved scalability for multi-node workloads and HPC workloads. Cloud RDMA is a networking infrastructure component that lets you build a true cloud HPC platform that can run scientific computations and ML/AI workloads. Cloud RDMA delivers price performance ratios comparable to on-premise infrastructure.
- H3 instances offer 88 vCPUs and 352 GB of DDR5 memory. H3 instances run on the Intel Sapphire Rapids CPU platform and Titanium offload processors. H3 instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance. H3 delivers performance improvements for a wide variety of HPC workloads such as molecular dynamics, computational geoscience, financial risk analysis, weather modeling, frontend and backend EDA, and computational fluid dynamics.
- C2 instances offer up to 60 vCPUs, 4 GB of memory per vCPU, and are available on the Intel Cascade Lake CPU platform. C2 instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance.
- C2D instances offer up to 112 vCPUs, up to 8 GB of memory per vCPU, and are available on the third generation AMD EPYC Milan platform. C2D instances are aligned with the underlying NUMA architecture to offer optimal, reliable, and consistent performance.
## Network-optimized machine family guide
The [**network-optimized machine family**](https://docs.cloud.google.com/compute/docs/network-optimized-machines)
is purpose-built for I/O-intensive workloads that require high network
bandwidth, superior packet processing performance, and exceptional block storage
throughput and IOPS per vCPU. C4N instances deliver the highest network and
block storage performance available in Compute Engine. C4N is an ideal
fit for workloads such as network and security appliances, high-performance
databases, high-scale data analytics and distributed file systems.
- C4N instances are available in eight predefined machine shapes across
`highcpu`, `standard`, and `highmem` configurations, ranging from 2 to 192
vCPUs and up to 1,488 GB of DDR5 memory.
Powered by the Intel Emerald Rapids CPU platform with a dual-NIC
[Titanium](https://cloud.google.com/titanium) offload architecture,
C4N instances are fully aligned with the underlying NUMA architecture to
deliver predictable, consistent performance.
C4N instances offer the following maximum performance capabilities:
- **Networking**: Up to 400 Gbps of network bandwidth and 95 million packets per second (MPPS).
- **Block storage**: Leading block storage performance with Hyperdisk Extreme, scaling up to 25 GiB/s of Hyperdisk bandwidth and up to 1,000,000 IOPS on the largest machine types.
- M4N instances offer up to 224 vCPUs, with up to 26.5 GB of memory per
vCPU, and are available on the Intel Emerald Rapids CPU platform. M4N
instances are aligned with the underlying NUMA architecture to offer
optimal, reliable, and consistent performance. The largest M4N machine types
come with two NICs to deliver the highest block storage performance and
network bandwidth. M4N delivers a new standard in storage performance,
doubling the current Hyperdisk Extreme performance compared to the M4
series, providing up to 25 GiB/s of aggregate performance and up to
1,000,000 IOPS on the largest machine types.
## Memory-optimized machine family guide
The [**memory-optimized machine family**](https://docs.cloud.google.com/compute/docs/memory-optimized-machines)
has machine series that are ideal for OLAP and OLTP SAP workloads, genomic
modeling, electronic design automation, and memory-intensive HPC
workloads. This family offers more memory per core than
any other machine family, with up to
48 TB
of memory.
- X5 bare metal instances offer up to 2,064 vCPUs, with 23.8 GB of memory per vCPU. X5 has machine types with 12, 16, 24, 32, and 48 TB of memory, and is available on the Intel Granite Rapids CPU platform.
- X4 bare metal instances offer up to 1,920 vCPUs, with either 12.8 or 17 GB of memory per vCPU. X4 has machine types with 6, 8, 12, 16, 24, and 32 TB of memory, and is available on the Intel Sapphire Rapids CPU platform.
- M4N dual-NIC instances offer up to 224 vCPUs, with up to 26.5 GB of memory per vCPU, and are available on the Intel Emerald Rapids CPU platform.
- M4 instances offer up to 224 vCPUs, with up to 26.5 GB of memory per vCPU, and are available on the Intel Emerald Rapids CPU platform.
- M3 instances offer up to 128 vCPUs, with up to 30.5 GB of memory per vCPU, and are available on the Intel Ice Lake CPU platform.
- M2 instances are available as 6 TB, 9 TB, and 12 TB machine types, and are available on the Intel Cascade Lake CPU platform.
- M1 instances offer up to 160 vCPUs, 14.9 GB to 24 GB of memory per vCPU, and are available on the Intel Skylake and Broadwell CPU platforms.
## Accelerator-optimized machine family guide
The [**accelerator-optimized machine family**](https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines)
is ideal for
[massively parallelized Compute Unified Device Architecture (CUDA) compute workloads](https://developer.nvidia.com/cuda-zone),
such as machine learning (ML) and high performance computing (HPC). This machine
family is the optimal choice for workloads that require accelerators
(GPUs or TPUs). This section provides information about GPU machines. For
information about machines that contain TPUs, see
[TPU machines](https://docs.cloud.google.com/compute/docs/tpus/tpu-machines).
Google also offers AI Hypercomputer for creating clusters of
accelerator-optimized VMs with inter-GPU communication, which are designed for
running very intensive AI and ML workloads. For more information, see
[AI Hypercomputer overview](https://docs.cloud.google.com/ai-hypercomputer/docs/overview).
### Arm
- A5X bare metal instances offer 176 vCPUs and 1,440 GB of memory. The A5X machine type has 4 NVIDIA Rubin GPUs attached to 2 NVIDIA Vera CPUs. A5X bare metal instances have a maximum network bandwidth of up to 7,200 Gbps.
- A4X Max bare metal instances offer up to 144 vCPUs and up to 960 GB of memory. Each A4X Max machine type has 4 NVIDIA B300 GPUs attached to 2 NVIDIA Grace CPUs. A4X Max bare metal instances have a maximum network bandwidth of up to 3,600 Gbps.
- A4X instances offer up to 140 vCPUs and up to 884 GB of memory. Each A4X machine type has 4 NVIDIA B200 GPUs attached to 2 NVIDIA Grace CPUs. A4X instances have a maximum network bandwidth of up to 2,000 Gbps.
> [!IMPORTANT]
> **Important:** The [Compute Engine Service Level Agreement (SLA)](https://cloud.google.com/compute/sla) doesn't apply to the A5X, A4X Max, and A4X machine types.
### x86
- A4 instances offer up to 224 vCPUs and up to 3,968 GB of memory. Each A4 machine type has 8 NVIDIA B200 GPUs attached. A4 instances have a maximum network bandwidth of up to 3,600 Gbps and are available on the Intel Emerald Rapids CPU platform.
- A3 instances offer up to 224 vCPUs and up to 2,952 GB of memory. Each A3 machine type has either 1, 2, 4, or 8 NVIDIA H100 or 8 H200 GPUs attached. A3 instances have a maximum network bandwidth of up to 3,200 Gbps and are available on the following CPU platforms:
- Intel Emerald Rapids - A3 Ultra
- Intel Sapphire Rapids - A3 Mega, High, and Edge
- A3 instances are available with the A3 Edge machine type (`a3-edgegpu-8g-nolssd`), which offers 208 vCPUs, 1,872 GB of memory, and 8 NVIDIA H100 GPUs, on the Intel Sapphire Rapids CPU platform and [Titanium](https://cloud.google.com/titanium).
- A2 instances offer 12 to 96 vCPUs, and up to 1,360 GB of memory. Each A2 machine type has either 1, 2, 4, 8, or 16 NVIDIA A100 GPUs attached. A2 instances have a maximum network bandwidth of up to 100 Gbps and are available on the Intel Cascade Lake CPU platform.
- G4 instances offer 6 to 384 vCPUs and up to 1,440 GB of memory. Each G4 instance has NVIDIA RTX PRO 6000 GPUs attached as follows:
- Whole GPUs in quantities of 1, 2, 4, or 8
- Fractional GPUs in quantities of 1/8, 1/4, or 1/2
- G4 instances have a maximum network bandwidth of up to 400 Gbps and are available on the AMD EPYC Turin CPU platform.
- G2 instances offer 4 to 96 vCPUs and up to 432 GB of memory. Each G2 machine type has either 1, 2, 4, or 8 NVIDIA L4 GPUs attached. G2 instances have a maximum network bandwidth of up to 100 Gbps and are available on the Intel Cascade Lake CPU platform.
## Machine series comparison
Use the following table to compare each machine family and determine
which one is appropriate for your workload. For a comparison of TPU versions,
see [TPU machines](https://docs.cloud.google.com/compute/docs/tpus/tpu-machines).
If, after reviewing this section, you are still unsure which family is best for
your workload, then start with the general-purpose machine family. For details
about all supported processors, see
[CPU platforms](https://docs.cloud.google.com/compute/docs/cpu-platforms).
To learn how your selection affects the performance of disk volumes
attached to your compute instances, see the following:
- Persistent Disk: [Disk performance by machine type and vCPU count](https://docs.cloud.google.com/compute/docs/disks/performance#machine-type-disk-limits)
- Google Cloud Hyperdisk: [Hyperdisk performance limits](https://docs.cloud.google.com/compute/docs/disks/hyperdisk-perf-limits#machine-type-perf)
Compare the characteristics of different machine series, from C4 to G2.
You can select specific properties in the
**Choose instance properties to compare** field to compare those properties
across all machine series in the following table.
Workload type Instance type CPU type Architecture vCPUs vCPU definition Memory Shared memory architecture Custom machine types Extended memory Sole tenancy Nested virtualization Confidential Computing Disk interface type Hyperdisk Balanced Hyperdisk Balanced HA Hyperdisk Extreme Hyperdisk ML Hyperdisk Throughput Local SSD Max Local SSD Standard PD Balanced PD SSD PD Extreme PD Network interfaces Network performance Tier_1 networking Multi-NIC Max GPUs Sustained use discounts Resource-based committed use discounts (CUDs) Compute flexible CUDs Spot VM discounts
## GPUs and compute instances
GPUs accelerate workloads on A5X, A4X Max, A4X, A4, A3, A2, G4, G2, and N1
instances. When you create an instance based on these machine series, except for
N1, Compute Engine automatically attaches the GPUs. For N1 instances,
you can attach GPUs during or after instance creation. No other machine series
support GPUs.
Accelerator-optimized instances have a fixed number of GPUs and vCPUs, and a fixed
amount of memory per machine type, with the exception of G2 machines that offer a
custom memory range. N1 instances with fewer GPUs attached are limited to a maximum
number of vCPUs. In general, a higher number of GPUs lets you create instances with a
higher number of vCPUs and more memory.
For more information, see [GPUs on Compute Engine](https://docs.cloud.google.com/compute/docs/gpus).
## What's next
- Learn how to [create and start a VM](https://docs.cloud.google.com/compute/docs/instances/create-start-instance).
- Learn how to [create a VM with a custom machine type](https://docs.cloud.google.com/compute/docs/instances/creating-instance-with-custom-machine-type).
- Complete the [Quickstart using a Linux VM](https://docs.cloud.google.com/compute/docs/create-linux-vm-instance).
- Complete the [Quickstart using a Windows VM](https://docs.cloud.google.com/compute/docs/create-windows-server-vm-instance).
- Learn more about attaching [block storage](https://docs.cloud.google.com/compute/docs/disks) to your VMs.