milvus-logo

GPU-enabled Milvus

Overview

Milvus is available in CPU-only and GPU-enabled distributions:

  • The CPU-only Milvus distribution relies on CPU exclusively to search and build indexes.
  • The GPU-enabled Milvus distribution supports GPU acceleration for search and index building. For example, CPU can be used for search while GPU is used for index building, improving query efficiency.

For GPUs that support CUDA, the GPU-enabled Milvus distribution can be used to achieve much better search performance when working with large-scale datasets.

Indexes for GPU-enabled Milvus

Milvus maps different embedding types with different index types. Click the tab below to view the index types supporting your embedding type.

Index type Indexing with CPU Indexing with GPU Search with CPU Search with GPU
FLAT N/A N/A ✔️ ✔️
IVF_FLAT ✔️ ✔️ ✔️ ✔️
IVF_SQ8 ✔️ ✔️ ✔️ ✔️
IVF_SQ8H ✔️ ✔️ ✔️ ✔️
IVF_PQ ✔️ ✔️ ✔️ ✔️
RNSG ✔️ ✔️
HNSW ✔️ ✔️
Annoy ✔️ ✔️
  • An index built with GPU is identical to built with CPU. The only difference is the time to build the index: GPU usually takes less time.
  • If top_k > 2048, Milvus switches from GPU search to CPU search.
  • If nprobe > 2048, Milvus switches from GPU search to CPU search.
Index type Indexing with CPU Indexing with GPU Search with CPU Search with GPU
FLAT N/A N/A ✔️
IVF_FLAT ✔️ ✔️
On this page