Oracle Cloud Infrastructure Documentation

Using GPUs

To work with GPUs in Data Science, the tenancy needs a limit for that shape and the shape must be available in the region you want to use it in.

If you're not an enterprise customer, you must create a service limit increase to use GPU because by default, your tenancy has zero limit.
Note

Service limit isn't the same as the shape capacity. The Data Science GPU limit is the maximum number of GPU your tenancy can use if the shape is available. If all the shapes in the region are in use, you might receive an out-of-capacity error even if you have Data Science limit.

Requesting a GPU

  1. Check if you have limit for the shape you want to use in Data Science.
  2. Open the Console.
  3. (Optional) Change the region if it's not set to the region in question.
  4. Click Governance & Administration.
  5. Under Tenancy Management, click Limits, Quotas and Usage.
  6. For Service, select Data Science.
  7. For Scope, select the region in question,
  8. For Resource, select the GPU shape you want to use.
  9. For Compartment, select the root compartment.
  10. Ensure the available count has enough limit.
  11. To increase the limit, click Request a service limit increase.

    Ensure the service category on the service limit request is set to Data Science.

    For more information on requesting a service limit increase, see the tutorial, Increase Your Data Science Service Limits.

Using Reserved GPU Capacity

Compute offers a capacity reservation feature giving tenants the ability to reserve GPU shapes. If you have reserved GPU capacity with Compute, you can transfer the GPUs to be used in Data Science. Create a support request ticket and specify the shape, count, region, tenancy, and reservation ID of the GPU reservation to be transferred to Data Science.
Important

At the moment, A10 capacity reservations can't be accepted, and A100 capacity reservations can only be accepted in certain regions.
When Data Science has completed the reservation, the tenancy is billed for the shape whether it's in use or not.
Note

After the reserved GPU is transferred to Data Science, Data Science begins billing for the GPU shape whether the GPU is used or not.

Tips for Using GPU with AI Quick Actions

If you have increased the service limit in Data Science to use a specific GPU shape for model deployment or fine tuning in AI Quick Actions, don't create a notebook session with that shape to access AI Quick Actions. If you do so, the GPU for the notebook session is counted toward the service limit, so the shape can't be used for model deployment or fine tuning. Select a different shape (CPU or GPU) to create a notebook to access AI Quick Actions.

Troubleshooting

If you meet any issues with using GPUs, see the GPUs troubleshooting section for help.