Knowhere 2.1.0 Beta
Pre-release
Release Note
Knowhere-2.1.0 Beta
We are happy to announce the release of Knowhere 2.1.0 Beta! Knowhere 2.1.0 newly supports RAFT GPU-based IVF index and FAISS GPU-based IVF index.
Special Thanks
@wphicks @cjnolet from NVIDIA contributed the RAFT integration code!
Breaking Change
- Support FAISS GPU-based IVF-series indexes. (#708)
- Support RAFT GPU-based IVF-series indexes. (#712, #734)
RAFT is an open-source library for GPU-based ANN search, developed by Nvidia. It offers impressive performance and a dependable GPU-based ANN index. With Knowhere 2.1.0, Nvidia has incorporated its advanced GPU solutions to enhance Knowhere's capabilities.
We conducted benchmark tests on RAFT's GPU-based IVF-series indexes and achieved remarkable results. All benchmarks are conducted in a host of 8c32g, and an Nvidia A100 GPU. NQ in these benchmarks are 100.
Benchmark1
First we did some comparison between RAFT-IVF-Flat with IVF-Flat and HNSW at 95% recall.
| DataSets | SIFT | GIST | GLOVE | DEEP |
|---|---|---|---|---|
| HNSW (VPS) | 14,537 | 791 | 1,516 | 5,761 |
| IVF-Flat (VPS) | 3,097 | 142 | 791 | 723 |
| RAFT-IVF-Flat (VPS) | 121,568 | 5,737 | 20,163 | 16,557 |
Benchmark2
Also we benchmarked RAFT-IVF-PQ comparing Knowhere's fastest index HNSW at 80% recall.
| DataSets | SIFT | GIST | GLOVE | DEEP |
|---|---|---|---|---|
| HNSW (VPS) | 20,809 | 2,593 | 8,005 | 13,291 |
| RAFT-IVF-PQ (VPS) | 271,885 | 7,448 | 38,989 | 80,363 |
Enhancement
- Improve performance of CPU-based IVF-series and BruteForce. (#703)
- Related Issue: milvus-io/milvus#22163
Bug fix
- Fix TANIMOTO distance calculation error in BruteFoce. (#719)
