Unlocking GPU Powers for Vector Search: OpenSearch's New Frontier
summary
Shows a drop‑in architecture that uploads segment vectors to an object store and offloads index building to a remote GPU build fleet, using NVIDIA CAGRA to produce a base graph that is converted via KAGRA→HNSW conversion into HNSW files CPUs can read. Claims 9–30x faster index builds, 2.5x lower CPU during ingest, ~3.4x cost savings, and 40–60% smaller on‑disk or memory footprints; uses segment size thresholds, fallbacks to CPU, and Docker images to deploy the service.
AI-written summary · about
chapters (18)
- 0:00GPU-powered vector search intro
- 1:27Vector index architecture
- 4:36Faster index builds matter
- 7:432024 indexing improvements
- 10:11CPU bottlenecks in indexing
- 13:03Benefits of 2024 optimizations
- 16:15GPU-based index acceleration
- 19:26Remote GPU fleet benefits
- 21:34Kagra algorithm for GPU indexing
- 24:47Searching GPU indices on CPU
- 29:02Tuning Kagra parameters
- 30:26Performance with GPU acceleration
- 31:26Enabling GPU indexing
- 35:00Configuring GPU acceleration
- 36:03Remote GPU index build setup
- 39:36Future GPU acceleration plans
- 41:21GPU features availability
- 42:26GPU fallback and stability
related talks
43mcuVS and Lucene: GPU-based Vector Search
37mMuves: Multimodal & multilingual vector search w/ Hardware Acceleration
29mUnlocking the Future of AI: Trends, Infrastructure, and Tokenomics
05mFrom Open Source to Real-World Scale
16mHyperbolic
41mScaling Welfare With Open Source: How NAV Migrated To Managed OpenSearch
13mDemocratizing AI with Open Source Compute
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing