What are compute targets in Azure Machine Learning?
A compute target is a designated compute resource or environment where you run your training script or host your service deployment. This location might be your local machine or a cloud-based compute resource. Using compute targets makes it easy for you to later change your compute environment without having to change your code.
Azure Machine Learning has varying support across different compute targets. In a typical model development lifecycle, you might:
- Start by developing and experimenting on a small amount of data. At this stage, use your local environment, such as a local computer or cloud-based virtual machine (VM), as your compute target.
- Scale up to larger data, or do distributed training by using one of these training compute targets.
- After your model is ready, deploy it to a web hosting environment with one of these deployment compute targets.
The compute resources you use for your compute targets are attached to a workspace. Compute resources other than the local machine are shared by users of the workspace.
Training compute targets
As you scale up your training on larger datasets or perform distributed training, use Azure Machine Learning compute to create a single- or multi-node cluster that autoscales each time you submit a job. You can also attach your own compute resource, although support for different scenarios might vary.
Compute targets can be reused from one training job to the next. For example, after you attach a remote VM to your workspace, you can reuse it for multiple jobs. For machine learning pipelines, use the appropriate pipeline step for each compute target.
You can use any of the following resources for a training compute target for most jobs. Not all resources can be used for automated machine learning, machine learning pipelines, or designer. Azure Databricks can be used as a training resource for local runs and machine learning pipelines, but not as a remote target for other training.
| Training targets | Automated machine learning | Machine learning pipelines | Azure Machine Learning designer |
|---|---|---|---|
| Local computer | Yes | ||
| Azure Machine Learning compute cluster | Yes | Yes | Yes |
| Azure Machine Learning serverless compute | Yes | Yes | Yes |
| Azure Machine Learning compute instance | Yes (through SDK) | Yes | Yes |
| Azure Machine Learning Kubernetes | Yes | Yes |
Tip
The compute instance has 120GB OS disk. If you run out of disk space, use the terminal to clear at least 1-2 GB before you stop or restart the compute instance.
Compute targets for inference
When performing inference, Azure Machine Learning creates a Docker container that hosts the model and associated resources needed to use it. This container is then used in a compute target.
The compute target you use to host your model affects the cost and availability of your deployed endpoint. Use this table to choose an appropriate compute target.
| Compute target | Used for | GPU support | Description |
|---|---|---|---|
| Azure Machine Learning endpoints | Real-time inference Batch inference |
Yes | Fully managed computes for real-time (managed online endpoints) and batch scoring (batch endpoints) on serverless compute. |
| Azure Machine Learning Kubernetes | Real-time inference Batch inference |
Yes | Run inferencing workloads on on-premises, cloud, and edge Kubernetes clusters. |
| Compute target | Used for | GPU support | Description |
|---|---|---|---|
| Local web service | Testing/debugging | Use for limited testing and troubleshooting. Hardware acceleration depends on use of libraries in the local system. | |
| Azure Machine Learning Kubernetes | Real-time inference | Yes | Run inferencing workloads in the cloud. |
| Azure Container Instances | Real-time inference Recommended for dev/test purposes only. |
Use for low-scale CPU-based workloads that require less than 48 GB of RAM. Doesn't require you to manage a cluster. Only suitable for models less than 1 GB in size. Supported in the designer. |
Note
When choosing a cluster SKU, first scale up and then scale out. Start with a machine that has 150% of the RAM your model requires, profile the result and find a machine that has the performance you need. Once you've learned that, increase the number of machines to fit your need for concurrent inference.
Azure Machine Learning compute (managed)
Azure Machine Learning creates and manages the managed compute resources. This type of compute is optimized for machine learning workloads. Azure Machine Learning compute clusters, serverless compute, and compute instances are the only managed computes.
There's no need to create serverless compute. You can create Azure Machine Learning compute instances or compute clusters from:
- Azure Machine Learning studio
- The Python SDK and the Azure CLI:
- An Azure Resource Manager template. For an example template, see Create an Azure Machine Learning compute cluster.
Note
Instead of creating a compute cluster, use serverless compute to offload compute lifecycle management to Azure Machine Learning.
When created, these compute resources are automatically part of your workspace, unlike other kinds of compute targets.
| Capability | Compute cluster | Compute instance |
|---|---|---|
| Single- or multi-node cluster | ✓ | Single node cluster |
| Autoscales each time you submit a job | ✓ | |
| Automatic cluster management and job scheduling | ✓ | ✓ |
| Support for both CPU and GPU resources | ✓ | ✓ |
Note
To avoid charges when the compute is idle:
- For a compute cluster, make sure the minimum number of nodes is set to 0, or use serverless compute.
- For a compute instance, enable idle shutdown. While stopping the compute instance stops the billing for compute hours, you'll still be billed for disk, public IP, and standard load balancer.
Supported VM series and sizes
Important
If your compute instance or compute clusters are based on any of these series, recreate with another VM size.
These series retired on August 31, 2024:
When you select a node size for a managed compute resource in Azure Machine Learning, you can choose from among select VM sizes available in Azure. Azure offers a range of sizes for Linux and Windows for different workloads. To learn more, see VM types and sizes.
There are a few exceptions and limitations to choosing a VM size:
- Some VM series aren't supported in Azure Machine Learning.
- Some VM series, such as GPUs and other special SKUs, might not initially appear in your list of available VMs. But you can still use them, once you request a quota change. For more information about requesting quotas, see Request quota and limit increases.
See the following table to learn more about supported series.
| Supported VM series | Category | Supported by |
|---|---|---|
| DDSv4 | General purpose | Compute clusters and instance |
| Dv2 | General purpose | Compute clusters and instance |
| Dv3 | General purpose | Compute clusters and instance |
| DSv2 | General purpose | Compute clusters and instance |
| DSv3 | General purpose | Compute clusters and instance |
| Ev3 | Memory optimized | Compute clusters and instance |
| ESv3 | Memory optimized | Compute clusters and instance |
| FSv2 | Compute optimized | Compute clusters and instance |
| M | Memory optimized | Compute clusters and instance |
| NCv3 | GPU | Compute clusters and instance |
While Azure Machine Learning supports these VM series, they might not be available in all Azure regions. To check whether VM series are available, see Products available by region.
Note
Azure Machine Learning doesn't support all VM sizes that Azure Compute supports. To list the available VM sizes, use the following method:
Note
Azure Machine Learning doesn't support all VM sizes that Azure Compute supports. To list the available VM sizes supported by specific compute VM types, use one of the following methods:
If you use the GPU-enabled compute targets, it's important to ensure that the correct CUDA drivers are installed in the training environment. Use the following table to determine the correct CUDA version to use:
| GPU Architecture | Azure VM series | Supported CUDA versions |
|---|---|---|
| Volta | NCv3 | 9.0+ |
In addition to ensuring the CUDA version and hardware are compatible, also ensure that the CUDA version is compatible with the version of the machine learning framework you're using:
- For PyTorch, you can check the compatibility by visiting Pytorch's previous versions page.
- For Tensorflow, you can check the compatibility by visiting Tensorflow's build from source page.
Compute isolation
Azure Machine Learning compute offers VM sizes that are isolated to a specific hardware type and dedicated to a single customer. Isolated VM sizes are best suited for workloads that require a high degree of isolation from other customers' workloads for reasons that include meeting compliance and regulatory requirements. Utilizing an isolated size guarantees that your VM is the only one running on that specific server instance.
The current isolated VM offerings include:
- Standard_M128ms
- Standard_F72s_v2
- Standard_NC24s_v3
- Standard_NC24rs_v3 (RDMA capable)
To learn more about isolation, see Isolation in the Azure public cloud.
Unmanaged compute
Azure Machine Learning doesn't manage an unmanaged compute target. You create this type of compute target outside Azure Machine Learning and then attach it to your workspace. Unmanaged compute resources can require extra steps for you to maintain or to improve performance for machine learning workloads.
Azure Machine Learning supports the following unmanaged compute types:
- Remote virtual machines
- Azure HDInsight
- Azure Databricks
- Azure Data Lake Analytics
For more information, see Manage compute resources.