ML Studio: Hardware Availability Limitations

Question

ML Studio: Hardware Availability Limitations

Arturo Sánchez Pineda 17

Hi team,

This is part of a collection https://learn.microsoft.com/en-ca/collections/rgru2toy67m56 from our company, dedicated to seeking support for issues and feedback related to our substantial usage of our private ML studio workspaces.

I must also say that all our usage is in Switzerland and North and West Europe Azure regions.

Issue Description:

Nvidia RTX GPUs are consistently unavailable in all compute cluster offerings, constraining the development and deployment of applications requiring specific GPU hardware like IsaacLab.

Suggested Solutions:

Expand Azure ML compute cluster hardware options to include Nvidia RTX GPUs or equivalent hardware, which is essential for advanced development requirements.

Thank you for your feedback and help,

Arturo

Prashanth Veeragoni 4,930 Reputation points Microsoft External Staff Moderator

2025-06-19T11:39:24.7166667+00:00

Hi Arturo Sánchez Pineda,

Thank you for your feedback. We understand your need for Nvidia RTX GPUs in Azure ML compute clusters to support workloads like IsaacLab.

Currently, RTX GPUs are not offered in Azure ML Studio in the Switzerland and Europe regions, but we suggest the following:

· Explore alternative GPU SKUs such as V100 (NCv3), A100 (NDv2), and T4 (NC T4 v3) that may meet your needs.

· Consider deploying a custom VM with RTX GPU (using N-Series VMs) and connecting it to your ML workspace.

Azure Machine Learning supports below VM series

Your suggestion is noted and will share the same feedback to internal product team.

Attaching documentation for your reference:

What are compute targets in Azure Machine Learning

Hope this helps.

Thank you!
Prashanth Veeragoni 4,930 Reputation points Microsoft External Staff Moderator

2025-06-20T12:57:11.33+00:00

Hi Arturo Sánchez Pineda,

Checking back to see if the above suggestion helped in resolving your issue, do let me know for further queries.

Thank you!
Prashanth Veeragoni 4,930 Reputation points Microsoft External Staff Moderator

2025-06-23T04:33:20.73+00:00

Hello Arturo Sánchez Pineda,

We haven't heard from you on the last response, if the above suggestion helped in resolving your issue please do up-vote or let us know for further queries.

Thank you!
Arturo Sánchez Pineda 17 Reputation points

2025-06-23T14:00:12.6566667+00:00
Hi @Prashanth Veeragoni, thanks for the reply!

.

Yes, we can see from our side:

.

Unfortunately, none of those are RTX and we need those for the research we are performing.

For reference, from https://docs.isaacsim.omniverse.nvidia.com/latest/installation/requirements.html

GPUs without RT Cores (A100, H100) are not supported.

.

So, as you rightly mentioned, get deployed dedicated VM with the needed type of GPU and attached to the workspaces

Consider deploying a custom VM with RTX GPU (using N-Series VMs) and connecting it to your ML workspace.
.
This still has some limitations and we needed to perform some hack in the SSH keys (please see previous question https://learn.microsoft.com/en-us/answers/questions/2265264)

.

So, if there is a way to call the attention to the team that administrate what type of machines are accessible via Compute Instances and/or Compute Clusters in ML Studio workspaces, will be really beneficial.

.

And I think for MS too, since those machines are quite expensive. It will be a win-win situation :)

.

Thanks for your suggestions and reeply,

Arturo
Arturo Sánchez Pineda 17 Reputation points

2025-06-24T08:57:31.4166667+00:00

Hi @Manas Mohanty ,Oh, I am sorry, I cut the link yesterday: https://learn.microsoft.com/en-us/answers/questions/2265264/i-need-to-attach-a-gpu-vm-created-in-the-same-vnet

My bad.

And yes, we were implementing almost one-to-one the solution you suggested (thanks), yet, as mentioned in the ticket, we need to make some customisations, and also we have some login differences regarding the machines that are accessible via Compute Instances.

I hope this time you can access the Question above.

Thanks again,

Arturo
Manas Mohanty 5,620 Reputation points Microsoft External Staff Moderator

2025-06-25T17:12:18.8033333+00:00

Hello Arturo Sánchez Pineda

Thank you for acknowledging that you are already implementing the proposed solution, Yes you mentioned you have to make update your SSH authentication.

I was thinking you could have opted for user/password authentication too.

Issue on mentioned ticket on more networking configuration of compute (compute with a public ip). We need to move the VM to Virtual network (to get a private ip) or whitelist the public ip in AML firewall and modify NSG to allow inbound connection from the public IP, but you mentioned you were able to connect through simply through a updated SSH.

Anyway, I think we are on same page for proposed solution. Please let us know consider using RTX Ubuntu images as a possible answer then as you have high priority set on RTX GPU card.

But Ideally N-series particularly ND24AS computes (costly ones though) are able to meet the requirement from lot of large language models in model catalogue. Some are event compatible with NC6 or NC24 computes which are possible alternate for any robust deep learning pipelines.

You can post a feature request about RTX GPU supported computes on feedback forum as they add new feature based on customer demand.

Thank you

1 answer

Your answer

Prashanth Veeragoni 4,930 Reputation points Microsoft External Staff Moderator

2025-06-19T11:39:24.7166667+00:00

Hi Arturo Sánchez Pineda,

Thank you for your feedback. We understand your need for Nvidia RTX GPUs in Azure ML compute clusters to support workloads like IsaacLab.

Currently, RTX GPUs are not offered in Azure ML Studio in the Switzerland and Europe regions, but we suggest the following:

· Explore alternative GPU SKUs such as V100 (NCv3), A100 (NDv2), and T4 (NC T4 v3) that may meet your needs.

· Consider deploying a custom VM with RTX GPU (using N-Series VMs) and connecting it to your ML workspace.

Azure Machine Learning supports below VM series

Your suggestion is noted and will share the same feedback to internal product team.

Attaching documentation for your reference:

What are compute targets in Azure Machine Learning

Hope this helps.

Thank you!
Prashanth Veeragoni 4,930 Reputation points Microsoft External Staff Moderator

2025-06-20T12:57:11.33+00:00

Hi Arturo Sánchez Pineda,

Checking back to see if the above suggestion helped in resolving your issue, do let me know for further queries.

Thank you!
Prashanth Veeragoni 4,930 Reputation points Microsoft External Staff Moderator

2025-06-23T04:33:20.73+00:00

Hello Arturo Sánchez Pineda,

We haven't heard from you on the last response, if the above suggestion helped in resolving your issue please do up-vote or let us know for further queries.

Thank you!
Arturo Sánchez Pineda 17 Reputation points

2025-06-23T14:00:12.6566667+00:00

Hi @Prashanth Veeragoni, thanks for the reply!

.

Yes, we can see from our side:

.

Unfortunately, none of those are RTX and we need those for the research we are performing.

For reference, from https://docs.isaacsim.omniverse.nvidia.com/latest/installation/requirements.html

GPUs without RT Cores (A100, H100) are not supported.

.

So, as you rightly mentioned, get deployed dedicated VM with the needed type of GPU and attached to the workspaces

Consider deploying a custom VM with RTX GPU (using N-Series VMs) and connecting it to your ML workspace.
.
This still has some limitations and we needed to perform some hack in the SSH keys (please see previous question https://learn.microsoft.com/en-us/answers/questions/2265264)

.

So, if there is a way to call the attention to the team that administrate what type of machines are accessible via Compute Instances and/or Compute Clusters in ML Studio workspaces, will be really beneficial.

.

And I think for MS too, since those machines are quite expensive. It will be a win-win situation :)

.

Thanks for your suggestions and reeply,

Arturo
Arturo Sánchez Pineda 17 Reputation points

2025-06-24T08:57:31.4166667+00:00

Hi @Manas Mohanty ,Oh, I am sorry, I cut the link yesterday: https://learn.microsoft.com/en-us/answers/questions/2265264/i-need-to-attach-a-gpu-vm-created-in-the-same-vnet

My bad.

And yes, we were implementing almost one-to-one the solution you suggested (thanks), yet, as mentioned in the ticket, we need to make some customisations, and also we have some login differences regarding the machines that are accessible via Compute Instances.

I hope this time you can access the Question above.

Thanks again,

Arturo
Manas Mohanty 5,620 Reputation points Microsoft External Staff Moderator

2025-06-25T17:12:18.8033333+00:00

Hello Arturo Sánchez Pineda

Thank you for acknowledging that you are already implementing the proposed solution, Yes you mentioned you have to make update your SSH authentication.

I was thinking you could have opted for user/password authentication too.

Issue on mentioned ticket on more networking configuration of compute (compute with a public ip). We need to move the VM to Virtual network (to get a private ip) or whitelist the public ip in AML firewall and modify NSG to allow inbound connection from the public IP, but you mentioned you were able to connect through simply through a updated SSH.

Anyway, I think we are on same page for proposed solution. Please let us know consider using RTX Ubuntu images as a possible answer then as you have high priority set on RTX GPU card.

But Ideally N-series particularly ND24AS computes (costly ones though) are able to meet the requirement from lot of large language models in model catalogue. Some are event compatible with NC6 or NC24 computes which are possible alternate for any robust deep learning pipelines.

You can post a feature request about RTX GPU supported computes on feedback forum as they add new feature based on customer demand.

Thank you

Answer 1

Hi Arturo Sánchez Pineda

You can search for NVIDIA RTX Virtual Workstation in Marketplace after click on "select all images" option while creating the VMs.

User's image

Those VMs can be attached with Attach compute option.

Could you share the full link https://learn.microsoft.com/en-us/answers/questions/2265264)

We can attach computes with username and password or ssh private key

User's image

https://learn.microsoft.com/en-us/azure/machine-learning/how-to-create-attach-compute-studio?view=azureml-api-2#other-compute-targets

Thank you

Share via

ML Studio: Hardware Availability Limitations

1 answer

Your answer