[Benchmark] qingming-engine Vector Search Performance: RX 7900 XTX 24G Shows Excellent Results

[patent rune]

SIFT-1M Results: QINGMING-ENGINE on AMD RX 7900 XTX achieves a throughput of 6,275.72 QPS with a P99 latency of 11.214 ms and a recall rate of 99.26%@1/100%@10 for searching 1 million 128-dimensional vectors.
GIST-10M Subset Results: For a 1-million subset of 960-dimensional vectors, the system maintains 470.29 QPS with a P99 latency of 25.755 ms and a high recall of 99.40%@1/100%@10.

The author has open-sourced it. Anyone want to give it a try?
https://github.com/uulong950/qingming-flat/blob/main/README.md

[signal raven]

Hey, nice work! 🔥 6K+ QPS on SIFT-1M with that recall rate is solid.

Cool to see more vector search stuff running on RDNA3 👀

[raw mulch]

sounds very impressive

[patent rune]

The previous one was a brute-force implementation. I also have an ANN version, which delivers explosive performance with a QPS of 480k.recall@1 97%,p99=2ms on this device😆

the gist 960