How to obtain a virtual machine with a GPU
This page is ARCHIVED. Our agreement with Microsoft to gain Azure virtual machines is over and you should explore our other GPU options for the department. These include our Department GPU Servers and the shared Dietrich GPU Cluster.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The following is the process to request access to a virtual machine (VM) with GPU on Azure
- Determine the Azure Service and VM Family size you are looking for and calculate the total CPU quantity needed. (For example, a Standard_NC6s_v3 has 1 GPU (Nvidia V100) but 6 CPU, so you would request 6 CPU)fff
- From within the subscription, start the CPU Quota increase request process and ensure the appropriate Azure service is specified. (Ex: Virtual Machine, Azure Machine Learning, Azure Lab Services)
- Work with Azure Support on the request.
- If Support denies the request, e-mail Carl or the PhD Computing TA and they will give you the email of the right point of contact. He will escalate your Support Incident number to our internal contact and work with Azure Support on getting the request approved.
- Requests that get approved are typically fulfilled within 48 hours of escalation.
Steps 1 and 2 are detailed in the figure below. Step 3 entails communicating via email with a Microsoft Support Engineer.
Please note: The NCasT4v3, ND A100 v4, NDM A100 v4, and NDv2 are constrained globally, and we are not able to support sponsorship requests for these SKUs for the foreseeable future. Please also keep in mind that due to constraints, the number of GPUs and the region you select may need to be adjusted depending on the demand in the region. In some instances, choosing a different region, VM Family, and/or different quantity of GPUs may yield a different result.
In the (usually rare) case that the deployment of the approved VM does not go through completely, note that there is a known issue that prevents the CUDA extension to be installed correctly. If this happens, you can install it manually by following this guide. Since this is a quite involved process, if you find yourself stuck we suggest to go to the PhD computing office hours and/or contact Carl.
