QingMing Engine (青冥) is a sophisticated, cross-platform brute-force vector search engine tailored for edge computing. The engine leverages advanced optimizations in memory bandwidth and cache residency, facilitating exceptional performance with zero index construction time. This enables both exact and near-exact recall capabilities for various applications.

Key Features

• High-Performance Architecture: Designed for multiple platforms, QingMing supports:
  • NVIDIA GPUs: Optimized for RTX 5090D and server GPUs, utilizing qingming.cu
  • AMD GPUs: Designed for Radeon RX 7900 XTX with HIP/ROCm support via qingming.cpp
  • Mobile Devices: Accelerated for Snapdragon and Apple Silicon using qingming-mobile.cpp

Performance Benchmarks

The engine has been thoroughly benchmarked across different hardware setups:

• NVIDIA RTX 5090D (24GB)

  • SIFT-1M (128-dim): Achieves 9354 QPS with 99.26% recall.
  • GIST-1M (960-dim): Delivers maximum throughput of 621.076 QPS.
  • Deep-10M (96-dim): Handles 10 million vectors with a recall of 99.960% at 1137.779 QPS.

• AMD Radeon RX 7900 XTX (24GB)

  • SIFT-1M (128-dim): Yields 6275.720 QPS at 99.260% recall.

• Mobile Performance (Snapdragon 8 Gen 5)

  • Achieves ~8 ms latency for brute-force searches on 100,000 128-dim vectors with negligible thermal impact.

Supported Data Sources

Various datasets are utilized, with primary data sourced from ANN-Benchmarks Datasets.

Licensing and Support

QingMing Engine offers a dual-licensing structure:

• QingMing-Mobile: Available for Android, iOS, automotive, IoT, and smart camera applications.
• QingMing-GPU: Designed for NVIDIA and AMD servers and workstations.

For inquiries, including volume licensing and enterprise solutions, contact via email at zhangxiaolong950@gmail.com.

QingMing Engine empowers developers and organizations with cutting-edge vector search technology, enhancing efficiencies across diverse platforms.