GTC 2020: Accelerating SPTAG Library on the GPUs for Approximate Nearest Neighborhood Search
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Accelerating SPTAG Library on the GPUs for Approximate Nearest Neighborhood Search
Murat Guney, NVIDIA
The Approximate Nearest Neighborhood (ANN) search algorithm is essential to many machine-learning applications. For instance, vector similarity search, multimedia search, and duplicate entry search all employ ANN for handling very large datasets efficiently. There are two main approaches to implementing ANN: space partitioning trees and locality sensitive hashing. Although hashing methods are accelerated on the GPUs (RAPIDS and FAISS), to our knowledge there is no GPU-accelerated space partitioning ANN algorithm in the literature. We'll present the GPU acceleration of the SPTAG library, where both space partition tree and neighborhood graph construction are accelerated on GPUs. We'll discuss the data structures and algorithms developed for efficient GPU implementation. Finally, we'll discuss the performance characteristics and compare the execution times against a single-socket GPU.