How we accelerated KNN search in Manticore: two-pass HNSW traversal, batch processing, and AVX-512

▤ Summary

The article explains how Manticore improved KNN search performance in its HNSW implementation. It focuses on engineering changes such as two-pass traversal, batch processing, prefetching, and AVX-512 optimizations. The team compares performance across CPU architectures and shows gains for different k values and oversampling settings. It also notes that the improvements are available in Manticore Search 27.1.5. The piece is useful for users evaluating vector search performance and hardware-dependent tuning.

▦ Linked Companies

Manticore Search

$1M to $5M