View available GPU resources
Inspect ResourceSlices and pools to see which GPU, MIG, and VFIO devices are allocatable on your cluster.
GPU workloads
Task-oriented walkthroughs for running GPU workloads.
Start with GPU allocation to choose a base resource type: full GPU, MIG, or VFIO passthrough. Then use the corresponding guide to request that resource.
Kubernetes DRA consumable capacity enables independent ResourceClaim objects to share a full GPU or MIG device with scheduler-managed capacity accounting.
Refer to Consumable capacity to configure memory-based, unlimited, or fixed-count shares.
Time-slicing and Multi-Process Service are optional in-claim sharing strategies. Fabric Manager partitioning is an optional topology layer for full-GPU and VFIO devices. Refer to Fabric Manager partitioning to select and activate one complete NVSwitch partition.
For example manifests, refer to
demo/
in the repository.
Request full GPUs
Request full GPUs with ResourceClaimTemplates and CEL selectors.
Time-slicing
Share a single GPU between multiple containers using CUDA time-slicing.
Share GPUs with consumable capacity
Share full GPUs and MIG devices across independent ResourceClaims with consumable capacity.
MIG (Multi-Instance GPU)
Allocate MIG slices to containers using static or dynamic MIG.
Fabric Manager partitioning
Constrain full-GPU and VFIO claims to partitions managed by NVIDIA Fabric Manager.
KubeVirt VFIO GPU passthrough
Configure the DRA driver for NVIDIA GPUs for KubeVirt VFIO passthrough.