Accessing the special GPU resources
As part of the Petascale Campus Initiative, six new compute hosts with Nvidia V100 General Purpose Graphical Processing Units (GPGPUs) have been deployed in the Melbourne Research Cloud. They replace our end-of-life K80 hosts, and provide far more flexibility than their predecessors. These GPUs are excellent for machine and deep learning, image rendering and analysis, and large matrix manipulation. They can also be used to accelerate graphic interfaces, especially those requiring real-time 3D and volumetric rendering.
Access to these GPUs; however, are still under CLOSED BETA. We plan to move the access to OPEN BETA in late 2021.
This is a very limited resource designed to be equivalent to operating a computer with RTX 3000 series GPU. The initial flavour available will have vGPU cores with 8GB of GPU frame buffer, 5 vCPU cores, 60GB of RAM and 480GB of ultra-fast NVMe ephemeral storage. Bigger flavours are available on request based on the requirements of your project. Once approved, you can migrate from the default instance.
For very large workloads requiring multiple GPUs, we still recommend using the GPGPU pool on the HPC service Spartan. It is optimised for parallel batch GPU computing, while MRC vGPU offering will only support a GPU operation on a single GPU.
There are two OS images available at time of writing:
- Windows 10 Enterprise Edition
- Ubuntu 20.04 LTS with default Ubuntu Desktop
If you are interested in participating and providing feedback on this resource and already has an existing project in MRC, you could submit a request to have the vGPU-enabled virtual machine created in your project.
How to launch your GPU instance
In the CLOSED BETA, a DevOps team member will launch the GPU on your behalf. Once we move to OPEN BETA phase, this section will include instruction on how to launch your GPU instance.
Setting up your Ubuntu 20.04 LTS Desktop GPU instance
The instructions below assumed your knowledge on managing/accessing your instances in MRC. If not, please follow to our tutorials here.
- Once your instance is provisioned,
sshinto your instance and set the default user
sudo passwd ubuntu
Skip this step in the CLOSED BETA. Create a new
RDPsecurity group to allow
RDPaccess. Go to your project dashboard, navigate to
+ Create Security Group> Set
Skip this step in the CLOSED BETA. In this new security group, click
+ Add Rule. Add the following rule:
- CIDR: 10.0.0.0/8
RDP into your instance and verify your access.
Setting up your Windows 10 Enterprise GPU instance
Follow instruction as provided in by the DevOps team. The process is still under active development by our DevOps team.