.png){kind=icon}
.png){kind=icon}
Return
BAGS: A bit-partitioned quantization-aware architecture for accelerating graph-based approximate nearest neighbor search
.png){kind=icon}
PRE
AI
S
J
M
S
DOI：10.1016/j.sysarc.2026.103897.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Approximate nearest neighbor search (ANNS) is a core technique for efficiently retrieving similar vectors in vector databases, and is widely used in recommendation systems and retrieval-augmented generation (RAG). Among ANNS methods, hierarchical navigable small world (HNSW) graphs have been extensively adopted due to their favorable recall–latency trade-off. However, HNSW suffers from a pronounced memory-bound bottleneck, primarily due to memory bandwidth saturation when accessing high-dimensional raw vectors during graph traversal. To mitigate this issue, recent approaches reduce vector dimensionality or precision through quantization, but such techniques inevitably incur accuracy degradation. To compensate for this loss, a re-ranking stage is commonly employed to precisely re-evaluate a small set of promising candidates, thereby restoring accuracy with limited additional computation. Nevertheless, this strategy introduces two key challenges: increased hardware resource consumption and additional memory capacity requirements.
Journal
IF:
4.1
Papers:
2.9K
Citations:
4.2K
Organization
Cited Papers