Yes, NVIDIA cuVS is fully compatible with Milvus GPU acceleration, providing advanced algorithms like CAGRA for 40x faster index building versus CPU methods.
cuVS Algorithm Support
Milvus integrates NVIDIA cuVS, which exposes GPU-native indexing algorithms (CAGRA, IVF-PQ, IVF-Flat) optimized for Blackwell’s tensor architecture. CAGRA, a GPU-native graph-based algorithm, delivers the fastest index build and search times on modern GPUs.
CPU-GPU Index Interoperability
cuVS indexes are interoperable between GPU and CPU. Milvus can build indexes on Blackwell GPU at extreme speed, then serve queries from CPU memory when latency permits. This flexibility enables hybrid deployments balancing cost and performance.
Benchmark Results
a full-text search engine’s integration of cuVS achieved 12x improvement in indexing throughput and 7x decrease in force-merge latency on GPUs. Milvus deployments running cuVS-based indexing see comparable gains—billion-element indexes build in minutes instead of hours.
Open-Source Integration
Both Milvus and cuVS are open-source, eliminating licensing overhead. Self-hosted Milvus operators gain access to production-grade GPU vector search without proprietary vendor lock-in.