GPU Index

Milvus supports various GPU index types to accelerate search performance and efficiency, especially in high-throughput, low-latency, and high-recall scenarios. This topic provides an overview of the GPU index types supported by Milvus, their suitable use cases, and performance characteristics. For information on building indexes with GPU, refer to Index with GPU.

GPU acceleration can greatly improve the search performance and efficiency of Milvus, especially for high-throughput, low-latency and high-recall scenarios, and is also very friendly to large nq batch search secnario.

Milvus' GPU support is contributed by Nvidia RAPIDS team. The following are the GPU index types currently supported by Milvus.

GPU_CAGRA

GPU_CAGRA is a graph-based index optimized for GPUs, which performs well on inference GPUs. It's best suited for situations with a small number of queries, where training GPUs with lower memory frequency may not yield optimal results.

• Index building parameters

  Parameter	Description	Default Value
  intermediate_graph_degree	Affects recall and build time by determining the graph's degree before pruning. Recommended values are 32 or 64.	128
  graph_degree	Affects search performance and recall by setting the graph's degree after pruning. A larger difference between these two degrees results in a longer build time. Its value must be smaller than the value of intermediate_graph_degree.	64
  build_algo	Selects the graph generation algorithm before pruning. Possible values: IVF_PQ: Offers higher quality but slower build time. NN_DESCENT: Provides a quicker build with potentially lower recall.	IVF_PQ
  cache_dataset_on_device	Decides whether to cache the original dataset in GPU memory. Possible values: "true": Caches the original dataset to enhance recall by refining search results. "false": Does not cache the original dataset to save memory.	"false"

• Search parameters

  Parameter	Description	Default Value
  itopk_size	Determines the size of intermediate results kept during the search. A larger value may improve recall at the expense of search performance. It should be at least equal to the final top-k (limit) value and is typically a power of 2 (e.g., 16, 32, 64, 128).	Empty
  search_width	Specifies the number of entry points into the CAGRA graph during the search. Increasing this value can enhance recall but may impact search performance.	Empty
  min_iterations / max_iterations	Controls the search iteration process. By default, they are set to 0, and CAGRA automatically determines the number of iterations based on itopk_size and search_width. Adjusting these values manually can help balance performance and accuracy.	0
  team_size	Specifies the number of CUDA threads used for calculating metric distance on the GPU. Common values are a power of 2 up to 32 (e.g. 2, 4, 8, 16, 32). It has a minor impact on search performance. The default value is 0, where Milvus automatically selects the team_size based on the vector dimension.	0

GPU_IVF_FLAT

Similar to IVF_FLAT, GPU_IVF_FLAT also divides vector data into nlist cluster units, and then compares distances between the target input vector and the center of each cluster. Depending on the number of clusters the system is set to query (nprobe), similarity search results are returned based on comparisons between the target input and the vectors in the most similar cluster(s) only — drastically reducing query time.

By adjusting nprobe, an ideal balance between accuracy and speed can be found for a given scenario. Results from the IVF_FLAT performance test demonstrate that query time increases sharply as both the number of target input vectors (nq), and the number of clusters to search (nprobe), increase.

GPU_IVF_FLAT is the most basic IVF index, and the encoded data stored in each unit is consistent with the original data.

When conducting searches, note that you can set the top-K up to 256 for any search against a GPU_IVF_FLAT-indexed collection.

• Index building parameters

  Parameter	Description	Range	Default Value
  nlist	Number of cluster units	[1, 65536]	128

• Search parameters

  • Common search

    Parameter	Description	Range	Default Value
    nprobe	Number of units to query	[1, nlist]	8

• Limits on search

  Parameter	Range
  top-K	<= 256

GPU_IVF_PQ

PQ (Product Quantization) uniformly decomposes the original high-dimensional vector space into Cartesian products of m low-dimensional vector spaces, and then quantizes the decomposed low-dimensional vector spaces. Instead of calculating the distances between the target vector and the center of all the units, product quantization enables the calculation of distances between the target vector and the clustering center of each low-dimensional space and greatly reduces the time complexity and space complexity of the algorithm.

IVF_PQ performs IVF index clustering before quantizing the product of vectors. Its index file is even smaller than IVF_SQ8, but it also causes a loss of accuracy during searching vectors.

• Index building parameters

  Parameter	Description	Range	Default Value
  nlist	Number of cluster units	[1, 65536]	128
  m	Number of factors of product quantization	dim mod m == 0	4
  nbits	[Optional] Number of bits in which each low-dimensional vector is stored.	[1, 16]	8

• Search parameters

  • Common search

    Parameter	Description	Range	Default Value
    nprobe	Number of units to query	[1, nlist]	8

• Limits on search

  Parameter	Range
  top-K	<= 1024

GPU_BRUTE_FORCE

GPU_BRUTE_FORCE is tailored for cases where extremely high recall is crucial, guaranteeing a recall of 1 by comparing each query with all vectors in the dataset. It only requires the metric type (metric_type) and top-k (limit) as index building and search parameters.

For GPU_BRUTE_FORCE, no addition index building parameters or search parameters are required.

Conclusion

Currently, Milvus loads all indexes into GPU memory for efficient search operations. The amount of data that can be loaded depends on the size of the GPU memory:

• GPU_CAGRA: Memory usage is approximately 1.8 times that of the original vector data.
• GPU_IVF_FLAT and GPU_BRUTE_FORCE: Requires memory equal to the size of the original data.
• GPU_IVF_PQ: Utilizes a smaller memory footprint, which depends on the compression parameter settings.