New applications for Nimbus GPU instances have been discontinued as of 29th February 2024.

Due to the high demand and limited availability of GPU nodes in Nimbus, we are limiting GPU access to fixed 6-month periods. This is independent of Nimbus projects themselves, which can be started at any time and renewed every 6 months (see https://support.pawsey.org.au/documentation/display/US/Apply+for+Nimbus+Cloud+access for details on how to apply for Nimbus access). If you are allocated one or more vGPU instances for a 6-month period, once the end of that period is reached, your project loses GPU access and your vGPU instance(s) will be deleted.

If you wish to continue using GPU access, you can re-join the end of the waiting list and wait for the next available 6-monthly allocation. You will know at the time of application which allocation period(s) you have been allocated (see below for details).

If there are more applications than available spaces, newer applications will get priority over existing allocations to ensure fair and equitable granting of resources.

You can request GPU resources when submitting an application for a Nimbus project via the application portal (https://apply.pawsey.org.au). If all you need are GPU resources, you do not need to request normal (non-GPU) resources to go with a new Nimbus project. If you have an existing Nimbus project that you wish to add GPU resources to, the project owner can add this via the application portal in the free text - project requirements field. Please contact help@pawsey.org.au if you have any issues.

NOTE: All CPU and GPU allocations should align so they both finish at the same time.

Allocation Periods

Each GPU allocation period lasts for 25 weeks and 3 days. The start and end dates for the next two years are:

Period	Apply By	Starts	Ends
S1 2023	Wed 21st Dec	Wed 4th Jan	Fri 30th Jun
S2 2023	Wed 21st Jun	Wed 5th Jul	Fri 29th Dec
S1 2024	Wed 20th Dec	Wed 3rd Jan	Fri 28th Jun
S2 2024	Wed 19th Jun	Wed 3rd Jul	Fri 27th Dec

If you wish to have GPU access for a specific period (or the next available period), you must apply on or before the application date for that period ("Apply By" in the above table). Availability will be based on the vGPU instance type you have requested (see below). You will then be able to create your vGPU on the start date of your allocated period. That access will finish at 5pm AWST on the end date for that same period. At that time, your vGPU instances will be shut down and deleted automatically, so you will need to make sure that any data or files you wish to keep are backed up either externally or to an attached storage volume within your Nimbus project before then. If you also have normal (non-vGPU) instances in your project, they will not be affected by this.

Requesting GPU Resources

There are two types of vGPU instances available:

Flavour	CPU Cores	RAM	Root Disk	vGPU
r1.V100-4C	7	45GB	80GB	4GB V100
r1.V100-8C	7	90GB	100GB	8GB V100

You can request up to two 4GB vGPUs or one 8GB vGPU in any one period for your Nimbus project. If additional vGPU instances are requested, additional discussion would be needed at the time of application to confirm the project's GPU requirements. Also, be aware that due to availability, larger vGPU allocations to a single project may need to be split across multiple allocation periods (for example, you request four 4GB vGPUs, but only two are available that period). In that situation, your vGPU instances will still be deleted at the end of each allocation period, as described above.

Deferring Access

You may decide to defer your GPU access until the next available allocation period. This could be because your initial allocation would otherwise be split across multiple periods, and you wish to delay so that all of your requested vGPU instances are available in the same period. If you wish to defer to a later allocation period, you must choose to do so at least 1 week before the application date for your allocated period, so that we can re-allocate your spot in that allocation period to another project.

Browser not supported