the resource name of "gpu.intel.com/xe" shows ambiguity

[zhm-algo]

the resource name of "gpu.intel.com/xe" shows ambiguity

we can distinguish which gpu it is

can we show "gpu.intel.com/b60" for Arc Pro B60, and "gpu.intel.com/b70" Arc Pro B70

[eero-t]

Use operator NFD option, so that nodes get labeled based on what type of GPUs are installed on given node. Then use suitable node selector in the workloads.

PS. following NFD rules are used to generate the labels: https://github.com/intel/intel-device-plugins-for-kubernetes/blob/main/deployments/nfd/overlays/node-feature-rules/platform-labeling-rules.yaml

[eero-t]

Note: same node should have only single type of GPUs, because K8s does not support anything else with device plugins.

For more advanced usage than this, you would need to use DRA GPU driver instead of GPU device plugin:

• https://kubernetes.io/docs/concepts/scheduling-eviction/dynamic-resource-allocation/
• https://github.com/intel/intel-resource-drivers-for-kubernetes/tree/main/doc/gpu#readme

[tkatila]

While naming the resources based on the gpu type is possible, I'd say it's not something we are eager to do. As @eero-t noted, typically a node has same GPU types. Then it's easier to keep the gpu resource intact and have a nodeSelector or nodeAffinity and place the Pod to certain node with certain resources (e.g. gpu pci id).

[eero-t]

@zhm-algo OK to close?