AI Training - Features, Capabilities and Limitations
AI Training is covered by OVHcloud Public Cloud Special Conditions.
Objective
This page provides the technical capabilities and limitations of the Public Cloud AI Training offer.
Features
Available features
AI Training allows you to train your models easily, with just a few clicks or commands. This solution runs your training job on the computational cloud resources you have chosen (CPU/GPU).
Logs and Monitoring tools
Command line interface (CLI)
AI Training is compliant with the OVHcloud AI CLI. Discover how to install the OVHcloud AI CLI.
Logs
To check the logs of your job, you can do it via the ovhai CLI using the following command:
Monitoring tools
To see information of your job, you can do so with the ovhai CLI using the command above:
You can then access your metrics through the Monitoring Url.
You are also able to check it from the OVHcloud Control Panel in your job information by clicking the Go to Graph Dashboard button.
Planned features
We continuously improve our offers. You can follow, vote and submit ideas to add to our roadmap at https://github.com/ovh/public-cloud-roadmap/projects/4.
Capabilities and limitations
Supported regions for jobs
AI Training can be used from any country in the world, as long as you have an OVHcloud account. Physically, two datacenters are available:
BHS(Beauharnois, Canada, America)GRA(Gravelines, France, Europe)
Attached resources
Compute resources
You can either choose the number of GPUs or CPUs for an AI Training job, not both.
By default, a job uses one GPU.
The memory resource is not customisable.
If you choose GPU:
- CPU, memory and local storage resources are not customisable but scaled linearly with each additional GPU.
If you choose CPU:
- Memory and local storage resource is not customisable but scaled linearly with each additional CPU.
The maximum amount of CPU/GPU, memory per CPU/GPU and local storage is available on the OVHcloud website, Control Panel and the ovhai CLI.
For your information, the current limits are:
- CPU: 12 per job.
- GPU: 4 per job.
Available hardware for AI Training
Currently, we provide:
- NVIDIA V100S (pricing available here).
Available storage
Local storage
Each AI Training job comes with a local storage space, which is ephemeral. When you delete your job, this storage space is also deleted. This storage space depends on the selected instances during your AI Training job creation. Please refer to the compute resources section for more information.
Local storage is limited and not the recommended way to handle data, see the OVHcloud documentation on data for more information.
Attached storage
You can attach data volumes from Public Cloud Object Storage. The Object Storage bucket should be in the same region as your AI Training job. Attached storage allows you to work on several TB of data, while being persistent when you delete your AI Training job.
Maximum execution time
There is no duration limitation on AI Training job execution.
Pre-installed AI environments
OVHcloud AI Training comes with pre-installed AI environments.
List of available AI Environments:
- AutoGluon + MXNet
- FastAI
- HuggingFace
- Miniconda (Python generic)
- MXNet
- One image to rule them all
- PyTorch
- Scikit-Learn
- TensorFlow
Environment customization
Each environment can be customized directly with PIP or CONDA (we support almost any package and library).
You can also use your own Docker images.
Docker images can be hosted in a public or private registry.
The use of docker-compose is not possible.
Please be aware that images need to be built with an AMD architecture.
Learn how to build and use your custom Docker image in this tutorial.
Network
-
Public networking can be used for all the AI Tools.
-
Private networking (OVHcloud vRack) is not supported.
Available ports to public network
Each job has a public URL, by default this URL accesses the port 8080 of the job. The default port can be configured when you submit a new job.
You can also access other ports by appending them to the URL.
Job URL for accessing the default port (starting with the job's ID):
Job URL for accessing the port 9000 (starting with the job's ID followed by the port number):
Only the HTTP layer is accessible.
Quotas per Public Cloud project
Each Public Cloud project grants a customer by default a maximum of 4 GPUs used simultaneously. Reach out to our support if you need to increase this limitation.
Go further
Browse the full AI Training documentation to further understand the main concepts and get started.
If you need training or technical assistance to implement our solutions, contact your sales representative or click on this link to get a quote and ask our Professional Services experts for a custom analysis of your project.
Feedback
We would love to help answer questions and appreciate any feedback you may have.
Are you on Discord? Connect to our channel and interact directly with the team that builds our AI services!