PiPNN removes the search step from graph index builds, cutting billion-vector build times from days to minutes

qdrant_engine · x · 2026-11-17

PiPNN is a new ultra-scalable graph-based nearest-neighbor indexing method that eliminates the search step from graph construction. Fast graph indexes like HNSW and Vamana spend most of their build time searching for neighbors, causing random memory access, cache misses, and idle CPU cores—billion-vector builds can take hours or days. Benchmarks in the attached plots show PiPNN builds (e.g., 947–3449s) are dramatically faster than HNSW (9600–12600s) and Vamana on BigANN-1B, DEEP-1B, OpenAI-2M and other datasets, while keeping competitive QPS-recall curves.

Original post →

More from Infra

Infra channel →