Jasper: GPU-Accelerated ANNS with Quantization and Dynamic Updates

Researchers have developed Jasper, a GPU-native approximate nearest neighbor search (ANNS) system that addresses three key limitations of current GPU-accelerated ANNS: lack of efficient batch updates for evolving datasets, absence of quantization techniques to reduce memory bandwidth strain from high-dimensional vectors, and difficulty overlapping compute and memory due to data-dependent memory accesses in greedy search. Jasper enables fast batch updates without rebuilding indices from scratch, introduces efficient quantization that reduces data movement without costly random memory accesses, and optimizes GPU execution to overlap compute and memory. The system is designed for real-world machine learning and information retrieval applications where datasets evolve over time. The work is described in arXiv paper 2601.07048.