PyNNDescent release cuts sparse kNN-graph build from hours to minutes
leland_mcinnes · x · 2026-10-28
A new PyNNDescent release brings massive performance gains: sparse-data kNN-graph build times drop from hours to a couple of minutes, and ANN index build times fall sharply. Benchmark plots on glove-100-angular and sift-128-euclidean show the new version beating hnswlib and the old 0.5.13 release by an order of magnitude across nearly the whole recall range.
Related event: PyNNDescent 0.7 Delivers Massive kNN Graph Speedups(4 posts)→
More from Infra
- 230 agent-run experiments on LLM inference: 8 lessons from DeepSeek 4.1 Flash tuning — jiayq · 2026-10-29
- Together's Post-Training Uses Per-GPU Pricing to Cut Agent Rollout Costs — togethercompute · 2026-10-29
- GitHub Unveils New Git Architecture Hitting 7,200 Writes/sec for AI Collaboration — mariorod1 · 2026-10-29
- Running a 10B-parameter model on a garbage GPU is the late-night thrill of choice — lxfater · 2026-10-29
- NVIDIA CTO Michael Kagan on Mellanox, NVLink and Planet-Scale Computing — BenBajarin · 2026-10-29
- Indie researcher publishes SGTE, a from-scratch alternative inference architecture to beat the scaling wall — ReportProfessional57 · 2026-10-29