NVIDIA cuVS is the world's fastest open-source library for vector similarity search and clustering on GPU. Power your RAG pipelines, recommender systems, and semantic search with 21× faster indexing and 29× higher throughput than CPU.
NVIDIA cuVS is an open-source library built on the CUDA software stack. It contains state-of-the-art implementations of approximate and exact nearest neighbor search, clustering, and dimensionality reduction — all optimized for GPU parallelism.
Each algorithm is performance-tuned for the latest NVIDIA GPU architectures, from Ampere to Hopper.
Benchmarks from official NVIDIA testing. GPU vs CPU across index build time, cost, throughput, and latency.
From genomics to e-commerce, cuVS powers the similarity search backbone of modern AI systems.
Install via conda or pip, and run your first GPU-accelerated ANN search with just a few lines of Python.
conda install -c rapidsai -c conda-forge cuvscagra.build() — index construction runs entirely on GPU. Hours become minutes.cagra.search() for approximate nearest neighbor queries. Achieve 29× higher throughput than CPU.import numpy as np
from cuvs.neighbors import cagra
from pylibraft.common import DeviceResources
# Create GPU resource handle
handle = DeviceResources()
# 1M vectors, 128-dimensional embeddings
dataset = np.random.random(
(1_000_000, 128)
).astype(np.float32)
# Build CAGRA index on GPU
index_params = cagra.IndexParams(
metric="sqeuclidean",
intermediate_graph_degree=64,
graph_degree=32,
)
index = cagra.build(index_params, dataset)
# Search: find top-10 neighbors for 1000 queries
queries = np.random.random(
(1000, 128)
).astype(np.float32)
search_params = cagra.SearchParams()
distances, neighbors = cagra.search(
search_params, index, queries, k=10
)
# neighbors.shape → (1000, 10)
print(f"Found {neighbors.shape} results")cuVS is integrated into the world's leading vector databases, search engines, and ML frameworks.
Own the exact-match premium domain for NVIDIA's fastest-growing GPU library.
Perfect for vector database companies, AI infrastructure teams, and NVIDIA ecosystem partners.
Also listed on Afternic · Sedo · Dan.com
Everything you need to get started with cuVS — from notebooks to research papers.