NVIDIA cuVS
GPU-accelerated vector search and clustering library from NVIDIA RAPIDS. Provides 8-12x faster index building and queries with multiple language support (C, C++, Python, Rust). This is an OSS library.
About this tool
Overview
cuVS is a GPU-accelerated library for vector search and clustering from NVIDIA RAPIDS. Enables databases to scale up and out for massive-scale vector search workloads, delivering unmatched speed through GPU acceleration.
Performance Benefits
Index Building
- 12x faster index building on GPU at 95% recall (vs. CPU)
- Massive speedups for large datasets
- Parallel processing on GPU architecture
Query Performance
- 8x lower search latencies at 95% recall
- Higher throughput at all recall levels
- Better latency characteristics
Integration Performance
- Faiss integration: 12x faster builds, 8x lower latency
- Milvus on GPU: Significant speedups with cuVS
- Production-grade performance improvements
Supported Algorithms
- IVF-PQ: Inverted File with Product Quantization
- IVF-Flat: Inverted File with flat vectors
- CAGRA: CUDA-Accelerated Graph-based ANN
- HNSW: Hierarchical Navigable Small World
- Brute Force: Exact search option
Multi-Language Support
- C: Native C API
- C++: Modern C++ interface
- Python: Python bindings
- Rust: Rust bindings
CPU-GPU Interoperability
- Build index on GPU, search on CPU
- Flexible deployment options
- Hybrid architectures supported
- Optimal resource utilization
Key Features
- Fully open source on GitHub
- Part of NVIDIA RAPIDS ecosystem
- Production-ready and battle-tested
- Continuous optimization and updates
- Enterprise support available
Use Cases
Real-Time Applications
- Machine learning workflows (hours to near real-time)
- Natural language processing at scale
- Hybrid search systems
- Anomaly detection
Database Integration
- Vector database acceleration
- In-database GPU processing
- Distributed search systems
- Cloud-native deployments
Integration
Faiss
- Enhanced GPU-accelerated Faiss
- Drop-in performance improvements
- Maintained compatibility
Milvus
- Native cuVS integration
- GPU-accelerated vector search
- Supercharged performance
Custom Applications
- Direct API access
- Embeddable in applications
- Flexible architecture
Technical Architecture
- Exploits parallel GPU architecture
- Optimized CUDA kernels
- Memory-efficient implementations
- Scalable to multiple GPUs
GPU Requirements
- NVIDIA GPU with CUDA support
- Compute Capability 7.0+ recommended
- Sufficient GPU memory for workload
- CUDA Toolkit installation
Installation
Conda (Recommended)
conda install -c rapidsai -c conda-forge cuvs
From Source
git clone https://github.com/rapidsai/cuvs
Documentation
- Comprehensive API documentation
- Example notebooks and tutorials
- Performance tuning guides
- Integration examples
RAPIDS Ecosystem
Part of NVIDIA RAPIDS suite:
- cuDF: GPU DataFrames
- cuML: GPU Machine Learning
- cuGraph: GPU Graph Analytics
- cuVS: GPU Vector Search
Performance Considerations
- Optimal for large-scale workloads
- GPU memory vs. dataset size
- Batch size optimization
- Multi-GPU scaling
Community and Support
- GitHub: rapidsai/cuvs
- RAPIDS community forums
- NVIDIA developer resources
- Enterprise support available
Pricing
Free and open-source under Apache 2.0 license. No licensing costs. GPU hardware required.
Loading more......
Information
Categories
Tags
Similar Products
6 result(s)cuVS is an open-source library from RAPIDS for fast, GPU-accelerated vector search, useful for building high-performance vector databases.
RAFT is a suite of GPU-accelerated libraries for data science, including support for vector search and similarity operations, often used in vector database scenarios.
Milvus is a mature, open-source vector database maintained by Zilliz, supporting large-scale similarity search with multiple indexing strategies and GPU acceleration. It includes variants such as Milvus Lite (lightweight version), Milvus Standalone (single-machine deployment), and Milvus Distributed (Kubernetes-based deployment for large scale).
Header-only C++/Python library for fast approximate nearest neighbor search implementing the HNSW algorithm. Used by Spotify and others, offers 10x speed increase over Annoy. This is an OSS library.
Pure-Rust implementation of HNSW algorithm for approximate nearest neighbor search. Decouples graph from vector storage for flexible deployment. Supports dense floating point and quantized int8 vectors. This is an OSS library.
Python implementation of Nearest Neighbor Descent for k-neighbor-graph construction and ANN search. Targets 80%-100% accuracy with fast performance and supports wide variety of distance metrics. This is an OSS library.