CS-PQ: Cache-Friendly SIMD Product Quantization for Large-Scale ANNS Index Construction
This work addresses the computational bottleneck of product quantization construction for large-scale approximate nearest neighbor search, a key problem in high-dimensional vector databases.
CS-PQ proposes a cache-friendly, SIMD-optimized product quantization framework for large-scale ANNS index construction, achieving up to 10.7x speedup over state-of-the-art CPU implementations without sacrificing accuracy.
Product Quantization (PQ) construction is deeply integrated into vector index construction for Approximate Nearest Neighbor Search (ANNS). The rapid growth in vector dimensionality and volume has significantly increased the computational cost of PQ. Existing GPU-based PQ accelerations are ill-suited for PQ construction due to its "one-to-one" execution pattern (one compute, one data load, i.e., data transfer overhead dominates). Although CPU-based solutions are prevalent, they are essentially general-purpose designs that fail to capture the intrinsic characteristics of PQ construction.In this paper, we propose CS-PQ, a Cache-friendly, SIMD-optimized PQ framework based on modern CPUs. CS-PQ introduces a vector-oriented SIMD paradigm that decouples quantization granularity from SIMD width by vectorizing across PQ centroids rather than subvector dimensions. It further restructures the execution pipeline to improve cache locality and reformulates PQ computation to eliminate redundant operations while preserving correctness. Experiments on large-scale datasets show that CS-PQ achieves up to 10.7 times speedup over state-of-the-art CPU-based PQ implementations without sacrificing ANNS accuracy.