GPU accelerated IVF-RaBitQ available in cuVS

[tfeher]

NVIDIA cuVS provides a growing set of GPU-accelerated algorithms for vector search. With the 26.08 release, cuVS adds a GPU-native implementation of IVF-RaBitQ, a new algorithm that brings a state-of-the-art quantization method to the library. This is a result of a collaboration between the Vector Database Research Group at Nanyang Technological University in Singapore and the cuVS team.

To handle large datasets efficiently, quantization methods are required. One of the promising approaches is RaBitQ, a quantization method developed by Jianyang Gao et al. in 2024. RaBitQ stands out because it does not require storing codebooks, and it comes with theoretically grounded bounds on quantization error. IVF-RaBitQ combines RaBitQ and the IVF index and delivers a strong balance between index compression, search quality and throughput.

The results of this collaboration demonstrate that IVF-RaBitQ can benefit greatly from GPU acceleration. Benchmarks also show IVF-RaBitQ delivering significant throughput improvements compared to IVF-PQ for large batch search. Detailed results are published in the paper from Jifan Shi et. al. “GPU-Native Approximate Nearest Neighbor Search with IVF-RaBitQ: Fast Index Build and Search”, which will be presented at VLDB 2026.

This fruitful collaboration between research and industry is a case study in how research ideas can move quickly into production. Rather than staying an academic prototype, the GPU native IVF-RaBitQ was engineered from the start to fit into the cuVS codebase and benefit from reusable computation primitives provided by cuVS. The source code is now openly available on github (https://github.com/NVIDIA/cuvs) and the compiled binaries are available through conda and pip.