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.
More from Infra
- tinygrad.js runs the real tinygrad in your browser on your GPU via WebGPU — CatAstro_Piyush · 2026-11-17
- Data Center Backlash Grows Loud, With Historical Echoes of 'A Nuisance Has No Rights' — justin_hart · 2026-11-17
- Constrained Decoding for Function Calling Cuts Endpoint Throughput by 17%-90% — remilouf · 2026-11-17
- Terray taps NVIDIA CUDA kernels to speed drug-design model TerraBind up to 7x — AllThingsApx · 2026-11-17
- Big Tech locks up 60% of DRAM supply it can't deploy yet — Reddit proposal for use-it-or-lose-it auctions — politefella0 · 2026-11-17
- NVIDIA expands CUDA-X science libraries: black hole analysis drops from a month to 8 minutes — AllThingsApx · 2026-11-17