cuVS

cuVS is an open-source library from NVIDIA for GPU-accelerated vector search and data clustering. The library is designed for workloads that need to search large collections of embeddings quickly. Typical use cases include semantic search, recommendation systems, and retrieval-augmented generation.

The library focuses on vector embeddings, which are numerical representations of data such as text, images, and audio. cuVS provides exact and approximate nearest neighbor search methods so the user can balance speed, memory use, and search quality for a specific workload.

cuVS includes indexing approaches such as IVF-Flat, IVF-PQ, and graph-based search. These methods are useful when the system must search very large datasets with low latency. Because cuVS runs on NVIDIA GPUs with CUDA, the library can reduce both query time and index build time compared with CPU-only workflows.

cuVS can fit into existing data and AI stacks. The library can be used alongside tools and platforms such as FAISS, Milvus, and Lucene. NVIDIA provides interfaces for several languages, including C, C++, Python, Rust, Java, and Go, which makes the library suitable for both systems programming and application development.

SwissGPU workstations provide dedicated NVIDIA GPUs running on bare-metal hardware, which makes them well suited for cuVS workloads. You can install cuVS as part of a CUDA-based software stack and run vector search pipelines, retrieval systems, or clustering jobs directly on the workstation. Because each machine provides exclusive access to GPU, CPU, memory, and storage resources, the system can use the full hardware capacity for large embedding datasets and low-latency search services.

Resources

SwissGPU