Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about PathWeaver, a novel multi-GPU framework designed to scale and accelerate graph-based Approximate Nearest Neighbor Search (ANNS) for large datasets in this 16-minute conference presentation from USENIX ATC '25. Discover how researchers from Seoul National University address the scalability limitations of existing GPU-based ANNS solutions that treat additional GPUs merely as memory extensions. Explore three key innovations: pipelining-based path extension that reduces redundant search iterations through GPU-to-GPU communication, ghost staging that uses representative datasets to identify optimal query starting points, and direction-guided selection that filters irrelevant points early to minimize memory accesses and distance computations. Understand how PathWeaver achieves significant performance improvements, delivering 3.24× geomean speedup and up to 5.30× speedup on 95% recall rate compared to state-of-the-art multi-GPU ANNS frameworks, making it valuable for applications in recommendation systems, natural language processing, and computer vision.
