QingMing Engine (青冥)

High-Performance On-Device Vector Search for the Edge
高性能端侧暴力向量搜索引擎

---

📖 Introduction / 简介

QingMing (青冥) is a cross-platform, brute-force vector search engine engineered for the edge. By optimizing for memory bandwidth and cache residency, it achieves exact (or near-exact) recall with zero index construction time.

青冥是一个专为边缘计算打造的跨平台暴力向量搜索引擎。通过极致优化内存带宽和缓存驻留，它在无需构建索引的情况下，实现了全精度的检索效果。

Supported Architectures / 支持架构:

• NVIDIA GPU: qingming.cu (Optimized for RTX 5090D / Server GPUs)
• AMD GPU: qingming.cpp (HIP/ROCm for 7900 XTX)
• Mobile NPU/CPU: qingming-mobile.cpp (NEON-accelerated for Snapdragon/Apple Silicon)

Data Source / 数据来源: ANN-Benchmarks Datasets

---

🚀 Quick Start / 快速启动

2.1 NVIDIA RTX 5090D v2 (24GB)

Environment: Ubuntu 24.04 + CUDA 12.8

Build / 编译

# Compiler Version: nvcc 12.8.93 (sm_120)
nvcc -O3 -arch=sm_120 qingming-flat.cu -o flat-hdf5 \
    -I/usr/include/hdf5/serial -L/usr/lib/x86_64-linux-gnu/hdf5/serial \
    -lhdf5_cpp -lhdf5

Benchmarks / 性能实测

1. SIFT-1M (128-dim)

Result: 9354 QPS @ ~5.5ms Latency (Batch=10k) | Recall@1: 99.26%

=======================================================
   QINGMING-ENGINE v1.0.0 PRO [GPU-FLAT]
   ADAPTER: NVIDIA RTX 5090D v2 (24GB)
=======================================================
Dataset Vectors : 1,000,000
Dimension       : 128
VRAM Usage      : 497.17 MB
-------------------------------------------------------
Recall@1        : 99.260 % (FP32 Precision)
Recall@10       : 100.000 %
-------------------------------------------------------
Max Throughput  : 9354.902 QPS
Latency P50     : 5.365 ms
Latency P99     : 5.573 ms
=======================================================

2. GIST-1M (960-dim)

Result: High-Dimensional Throughput Test

Dataset Vectors : 1,000,000
Dimension       : 960
VRAM Usage      : 3702.74 MB
-------------------------------------------------------
Recall@1        : 99.400 %
Recall@10       : 100.000 %
-------------------------------------------------------
Max Throughput  : 621.076 QPS
Latency P99     : 1.186 ms
=======================================================

3. Deep-10M (96-dim)

Result: 10 Million Vectors Flat Search

Dataset Vectors : 9,990,000
Dimension       : 96
VRAM Usage      : 3666.12 MB
-------------------------------------------------------
Recall@1        : 99.960 %
Recall@10       : 99.990 %
-------------------------------------------------------
Max Throughput  : 1137.779 QPS
Latency P99     : 1.188 ms
=======================================================

---

2.2 AMD Radeon RX 7900 XTX (24GB)

Environment: Ubuntu 24.04 + ROCm 6.2

Build / 编译

/opt/rocm-6.2.4/bin/amdclang++ -x hip -O3 --offload-arch=gfx1100 qingming.cpp -o qingming_amd \
    -I/usr/include/hdf5/serial -L/usr/lib/x86_64-linux-gnu/hdf5/serial \
    -lhdf5_cpp -lhdf5

Benchmarks / 性能实测

SIFT-1M (128-dim)

=======================================================
QINGMING-ENGINE v1.0.0 PRO [REDMOON]
PLATFORM: AMD Radeon RX 7900 XTX (24GB)
Dataset Vectors : 1,000,000
Dimension       : 128
Recall@1        : 99.260 %
Max Throughput  : 6275.720 QPS
Latency P99     : 11.214 ms
=======================================================

GIST-1M (960-dim)

Dataset Vectors : 1,000,000
Dimension       : 960
Recall@1        : 99.400 %
Max Throughput  : 470.285 QPS
Latency P99     : 25.755 ms
=======================================================

---

2.3 Mobile: Xiaomi 17 Pro Max (Snapdragon 8 Gen 5)

Core Tech: NEON SIMD + L3 Cache Residency (Zero-Copy)

Build / 编译

$TOOLCHAIN/aarch64-linux-android34-clang++ -O3 -static-libstdc++ -flto \
    -march=armv8.2-a+fp16+dotprod qingming-mobile.cpp -o qingming_8gen5

Deployment / 部署

adb push qingming_8gen5 /data/local/tmp/
adb shell chmod +x /data/local/tmp/qingming_8gen5
adb shell /data/local/tmp/qingming_8gen5 100000

Performance / 性能表现

Latency: Brute-force search over 100k 128-dim vectors achieves ~8ms latency per query.
Power: 1000 consecutive queries resulted in negligible thermal increase.

---

💼 Licensing & Enterprise Support / 商业授权

QingMing Engine offers a dual-licensing model optimized for Edge and Server deployments. 青冥引擎提供针对边缘端和服务器端的双轨授权模式。

📱 QingMing-Mobile

Target: Android, iOS, Automotive (Smart Cockpit), IoT, Smart Cameras.
Tech: ARM CPU + NEON Optimization.

Region	Pricing Strategy	Terms
Global	$9.99 / device / year	Converts to perpetual license after 3 years. Or $19.99 one-time perpetual.
China	¥9.99 / 设备 / 年	3 年后自动转为永久授权。 或 ¥19.99 一次性永久授权。

💻 QingMing-GPU

Target: NVIDIA / AMD Servers, Workstations.
Tech: CUDA / HIP.

Region	Pricing Strategy	Terms
Global	$99 / GPU / year	Or $199 one-time perpetual.
China	¥99 / GPU / 年	或 ¥199 一次性永久授权。

Note: Vehicles/IoT devices using the CPU-based qingming-mobile.cpp fall under Mobile Pricing, even if a GPU is present.

---

✉️ Contact / 联系我们

Email: zhangxiaolong950@gmail.com

Volume licensing, source code access, and custom enterprise integration available upon request.
支持批量授权、源码交付与企业级定制方案。