milvus-operator icon indicating copy to clipboard operation
milvus-operator copied to clipboard

GPU Deployment in Milvus 2.4

Open qchenzi opened this issue 11 months ago • 3 comments

I am considering deploying Milvus 2.4 in K8S and would like to confirm the feasibility of the following configuration:

  • GPU Nodes: Deploying only the IndexNode and QueryNode components on GPU-equipped nodes.
  • Non-GPU Nodes: Deploying the remaining components—DataNode, MixCoord, Proxy—on nodes without GPU capabilities.

This setup aims to leverage GPU acceleration for indexing and query processing while utilizing non-GPU nodes for other tasks. Could you please confirm if this configuration is supported in Milvus.

Thank you for your assistance.

qchenzi avatar Feb 11 '25 02:02 qchenzi

Yes, it's supported. By the way, 2 things to be noted:

  1. Check our prerequisite guides for setup in https://milvus.io/docs/prerequisite-gpu.md.
  2. Use GPU Index Types, otherwise Milvus will still use CPU for Indexing https://milvus.io/docs/gpu_index.md

haorenfsa avatar Feb 11 '25 03:02 haorenfsa

  • [ ] Add docs for GPU deployment using milvus-operator

haorenfsa avatar Feb 11 '25 03:02 haorenfsa

Thank you for confirming that the configuration is supported.

qchenzi avatar Feb 17 '25 03:02 qchenzi