About Flexible GPUs

Flexible GPUs (fGPUs) are resources used for graphic rendering and large parallel computation in the Cloud. You can allocate fGPUs to your account and attach them to or detach them from your virtual machines (VMs).

General Information

An fGPU is a resource that can be allocated to your account, and attached to or detached from your VMs. You can attach 2 fGPUs to the same VM, provided that the fGPUs are the same models.

An fGPU can be in these different states:

  • allocated: The fGPU is reserved for your account.

  • attaching: The fGPU is scheduled for attachment to the specified VM. You need to stop the VM for the fGPU to reach the attached state.

  • attached: The fGPU is used by the VM. It is considered a device connected to the VM.

  • detaching: The fGPU is scheduled for detachment from its VM. You need to stop the VM for the fGPU to reach the allocated state.

fGPU Lifecyle

sch FCU fGPULifecycle

When you attach an fGPU, the VM must meet the following conditions (see the table of models of fGPUs below):

  • Have the highest (1) performance flag

  • Have less than a certain number of vCores

  • Have less than a certain amount of memory

  • Have a processor generation that is compatible with the fGPU

In addition, the VM must not be a dedicated VM.

Models of fGPUs

3DS OUTSCALE provides different models of fGPUs, with different amounts of video RAM (VRAM):

Model of fGPU VRAM of the fGPU (in MiB) Maximum Number of vCores Maximum Amount of Memory (in GiB) Compatible Processor Generations Regions

nvidia-a100

40000

35

250

v5, v6

  • eu-west-2

nvidia-a100-80

80000

35

256

v6

  • eu-west-2

  • cloudgouv-eu-west-1

nvidia-k2

4096

80

512

v3, v4

  • eu-west-2

  • us-east-2

  • us-west-1

nvidia-l40

48000

35

250

v7

  • eu-west-2

nvidia-m60

16000

80

512

v3, v4

  • eu-west-2

nvidia-p6

16000

80

512

v5

  • eu-west-2

  • cloudgouv-eu-west-1

  • us-east-2

  • us-west-1

  • ap-northeast-1

nvidia-p100

16000

80

512

v5

  • eu-west-2

  • us-east-2

  • us-west-1

nvidia-v100

16000

35

250

v5

  • eu-west-2

To make sure you have the most up-to-date values, use the ReadFlexibleGpuCatalog method as reference.

VM Type Conversion

As with all VMs, you can modify the type of a VM that has fGPUs attached to it. However, the number of attached fGPUs might be incompatible with the new VM type (number of vCores, amount of memory, and processor generation). Different behaviors occur depending on the conversion:

Old Type New Type Conversion Behavior

AWS

AWS

All fGPUs attached to the VM are deleted.

Then as many as possible are created and attached to the converted VM.

AWS

TINA

All fGPUs attached to the VM are deleted.

TINA

AWS

All fGPUs attached to the VM are detached.

Then as many as possible are created and attached to the converted VM.

TINA

TINA

If all fGPUs attached to the VM are compatible with the new VM type, the VM is converted.

If not, a vm-specs-mismatch error is returned.

For more information on modifying VM types, see Modifying a VM Attribute.

Related Pages

Corresponding API Methods

AWS™ and Amazon Web Services™ are trademarks of Amazon Technologies, Inc or its affiliates in the United States and/or other countries.

Export to PDF