SKS: GPU MIG partitioning support

Exoscale SKS now supports GPU MIG partitioning for MIG-capable GPU nodepools.

With Multi-Instance GPU (MIG), customers can split a supported physical GPU into multiple isolated GPU instances, each with dedicated compute and memory resources. This makes it easier to run several smaller GPU workloads on the same SKS node instead of assigning a full GPU to each workload.

Availability and setup

• MIG partitioning is available for Exoscale A30, RTX Pro 6000 and B300 GPUs.
• Standard SKS nodepools can be configured with MIG profiles through the Exoscale Portal, CLI, or Terraform provider.
• Karpenter-managed GPU nodepools can use MIG through the ExoscaleNodeClass configuration.

Why use MIG?

• Run multiple smaller GPU workloads on one physical GPU.
• Improve GPU utilization and reduce underused capacity.
• Keep workloads isolated at the GPU partition level.
• Better accommodate AI inference, development and testing, and smaller batch workloads.

Note: Existing nodes must be cycled before a changed MIG profile takes effect. Billing is based on the selected GPU node instance type, not on individual MIG partitions.

For setup instructions, see:

Configure GPU MIG partitioning on SKS