Accelerating OpenSearch K-NN With FAISS, Leveraging AVX-512 for a...
summary
FAISS is integrated into OpenSearch K-NN with new AVX-512 code paths that vectorize Hamming popcount and inner-product/L2 kernels, cutting inner-product cycle counts (example: 28.86 → 7.32 cycles) and delivering ~9–11% higher indexing throughput and ~7–10% better search throughput with ~6–10% lower latency on 768-d random datasets, including SQ16 quantized vectors; feature is upstreamed and enabled in OpenSearch 2.9.19 and benefits Intel 4th‑gen Xeon (C7i/M7i/R7i) instances. AVX-512 vectorized popcount FAISS + OpenSearch K-NN
AI-written summary · about
chapters (13)
- 0:00Vector search and hardware acceleration
- 3:12Agenda and objectives
- 11:57Scaling vector databases challenges
- 18:49SIMD and AVX-512 fundamentals
- 21:59SIMD vs multithreading benefits
- 23:46FAISS binary vectors and Hamming distance
- 27:29AVX-512 pop count for Hamming distance
- 30:22FAISS performance with AVX-512
- 32:36AVX-512 vs AVX-2 in FAISS
- 36:44Business impact and TCO
- 39:15Intel's OpenSearch and FAISS contributions
- 42:30Future optimizations and community
- 43:39Q&A
related talks
33mBeyond Benchmarks: Real-World AI Search Performance in OpenSearch (Lucene Vs Faiss Vs JVector)
55mVector Search Made Simple: Getting Started with OpenSearch for AI Applications
43mOpenSearch and the Path to a Trillion-scale Vector Database
48mUnlocking Next-gen Semantic Search With OpenSearch 3.x
26mOpenSearch and the Path to a Trillion-s...
40mUsing Native CPU Vectorized Instructions for 60% Speedup on Lucene's KNN Vector Search
32mUnder the Hood: Key Algorithms Powering Vector Search
30mVector Search Made Simple: Getting Started with OpenSearch for AI Applications